Trading with intelligence

How AI shapes and is shaped

by international trade
Disclaimer
This publication has been prepared under the WTO Secretariat’s own responsibility. It does not
necessarily reflect the positions or opinions of WTO members and it is without prejudice to their rights
and obligations under the WTO agreements. The opinions expressed and arguments employed herein
are not intended to provide any authoritative or legal interpretation of provisions of the WTO agreements
and shall in no way be read or understood to have any legal implications. The terms and illustrations used
in this publication do not constitute or imply an expression of opinion by the WTO Secretariat concerning
the status or boundaries of any territory. The opinion pieces written by the external contributors are the
sole responsibility of their respective authors.
Table of contents
Acknowledgements 02
Abbreviations 03
Foreword 04
Executive summary 06

Chapter 1 Introduction 11

Chapter 2 Why is AI a trade issue? 13

(a) W
 hat is artificial intelligence (AI) and what makes it unique? 14
(b) How will AI affect international trade? 18

Chapter 3 The policies of AI and trade 35

(a) AI and trade: key policy considerations 36
(b) T
 he global race to promote and regulate
AI and the risk of fragmentation 46

Chapter 4 What role for the WTO? 65

(a) Promoting global convergence 66
(b) F
 acilitating trade in AI-related goods and
services for AI growth and development 70
(c) Minimizing negative international spillovers 77
(d) H
 elping to address and prevent trade tensions
and frictions 79
(e) P
 romoting inclusiveness through special and
differential treatment and technical assistance 80
(f) S
 cholars’ views on the possible implications of AI
for international trade rules 82

Chapter 5 Conclusion 89

Annex 1 Key AI-related terms 93

Annex 2 Technical appendix on the simulation scenarios 94
Annex 3 Key policy and standard-related
international initiatives in the area of AI 98
Annex 4 Survey of academics 104
Bibliography 105
Opinion pieces
Richard Baldwin, “AI means that services will be the future of trade” 25
James Manyika, “Harnessing technology to advance shared prosperity” 37
Shin-yi Peng, “AI: Amplifying the digital trade issues” 41
Eduardo Paranhos, “Navigating AI regulation:
balancing innovation, risks and regulatory defragmentation” 45
Daniel Trefler, “Building global chains of trust” 81

1
 Acknowledgements
The report was prepared under the general responsibility and Division (Dolores Halloran, Suja Rishikesh Mavroidis,
guidance of Johanna Hill, WTO Deputy Director-General, Roberta Lascari, Xiaodong Wang); Trade and Environment
and Ralph Ossa, Director of the Economic Research and Division (Erik Wijkström, Devin McDaniels); Trade in Services
Statistics Division. and Investment Division (Pamela Apaza Lanyi, Antonia
Carzaniga, Xiaolin Chai, Markus Jelitto, Joscelyn Magdeleine,
Coordination of the report and preparation of the chapters Ruosi Zhang); Trade Policies Review Division (Rohini Acharya);
was led by Emmanuelle Ganne and Ankai Xu. The main and Rules Division (Clarisse Morgan, Hiromi Yano).
authors of the report are Emmanuelle Ganne, Lauro Locks and
Ankai Xu. Opinion pieces were provided by Richard Baldwin, James
Manyika, Eduardo Paranhos, Shin-Yi Peng and Daniel Trefler.
Substantial contributions were received from Eddy Bekkers, A case study was provided by the Ministry of Digital
Şeref Gökay Coşkun, Hryhorii Kalachyhin, Anastasiia Development and Information of Singapore.
Koltunova, Kathryn Lundquist, Martin Roy and Xiaoping Wu.
Additional contributions were received from Milena Azevedo, The following individuals from outside the WTO Secretariat
Arti Gobind Daswani, Jesse Nicol, Daniel Ramos, Yves provided useful comments during the initial drafting stage
Renouf, Stefania Semenova, Monia Snoussi Mimouni and of the report: Susan Aaronson, Andrea Andrenelli, Craig
Astghik Solomonyan. Tomasz Koziel also provided valuable Atkinson, Richard Baldwin, Erik Brynjolfsson, Mira Burri,
advice and support. Dan Ciuriak, Avi Goldfarb, Chiara del Giovane, Gael
Grooby, Janos Ferencz, Martina Ferracane, Emily Jones,
Valuable research assistance was provided by Renate Olia Kanevskaia, Kholofelo Kugler, Javier López González,
Busstra. Additional research support was provided by Heidi Lund, Jiabin Luo, Petros Mavroidis, Joshua Meltzer,
Prabisha Basnet, Marco Cheang, Saniya Khanna and Julia Neha Mishra, Hildegunn Kyvik Nordås, Eduardo Paranhos,
Collado Serrano. Frank Pasquale, Shin-Yi Peng, Robert Teh, María del
Carmen Vásquez Callo, Daniel Trefler and Sujin Yuk.
The following divisions in the WTO Secretariat provided
valuable comments on drafts of the report: Responses to the survey mentioned in Annex 4 were
received from Susan Aaronson, Dan Ciuriak, Johannes
Agriculture and Commodities Division (Rolando Alcala, Cédric Fritz, Olia Kanevskaia, Kholofelo Kugler, Heidi Lund,
Pene, Christiane Wolff); Council and Trade Negotiations Petros Mavroidis, Hildegunn Kyvik Nordås, Eduardo
Committee Division (Stefania Gallo); Development Division Paranhos and Shin-Yi Peng.
(Lucas Chiodi, Michael Roberts, Raúl Torres); Economic
Research and Statistics Division (Marc Auboin, Marc The production of the report was managed by Anthony
Bacchetta, Gabrielle Marceau, Jose-Antonio Monteiro, Roberta Martin and Serge Marin-Pache of the Information and
Piermartini); Legal Affairs Division (Jorge Castro, Maria Pereyra, External Relations Division. Helen Swain edited the report.
Muhammad Ahsan Ali); Intellectual Property, Government Gratitude is also due to the translators in the Languages,
Procurement and Competition Division (Wolf Meier-Ewert, Documentation and Information Management Division
Anna Caroline Müller, Antony Taubman); Market Access for the high quality of their work.

2
Abbreviations
AfCFTA African Continental Free special and
S&DT 
Trade Area differential treatment
AI artificial intelligence SDG United Nations Sustainable
Development Goal
CPU central processing unit
subsidies and
SCM 
Dispute Settlement
DSU 
countervailing measures
Understanding
sanitary and
SPS 
ECJ European Court of Justice
phytosanitary measures
EU AIA EU Artificial Intelligence Act
STC specific trade concern
GATS General Agreement on
TBT technical barriers to trade
Trade in Services
TPR trade policy review
GATT General Agreement on
Tariffs and Trade Trade Policy
TPRM 
Review Mechanism
generative artificial
GenAI 
intelligence trade-related
TRIMS 
investment measures
Government Procurement
GPA 
Agreement TRIPS trade-related aspects of
intellectual property rights
GPU graphics processing unit
UNCTAD UN Trade and Development
ICT information and communications
technology United Nations
UNDP 
Development Programme
International Electrotechnical
IEC 
Commission UNECE United Nations Economic
Commission for Europe
IMF International Monetary Fund
UNESCO United Nations
IoT Internet of Things
Educational, Scientific and
IP intellectual property Cultural Organization

International Organization
ISO  UNIDO United Nations Industrial
for Standardization Development Organization

ISO/IEC JTC/SC ISO/IEC Joint Technical WCO World Customs Organization

Committee and subcommittee
WTO World Trade Organization
IT information technology
Information Technology
ITA 
Agreement
International
ITU 
Telecommunication Union
LDC least-developed country
LLM large language model
MFN most-favoured nation
MSME micro, small and
medium-sized enterprise
NTM non-tariff measure
Organisation for
OECD 
Economic Co-operation
and Development
R&D research and development
RTA regional trade agreement

3
 Foreword
I am delighted to present the World Trade Organization
Secretariat’s first comprehensive report on artificial
intelligence (AI) and international trade. This report marks
a milestone in our efforts to understand the impacts AI is
having, and will continue to have, on global trade. As AI
continues to evolve and transform the ways we work,
live and do business, the global trade community must
recognize these impacts and respond to maximize the
gains for people, businesses and economies, and minimize
potential risks.

Declaring AI to be “the new electricity” has already become

a cliché, but it may still be an understatement. AI is a
general-purpose technology that has entered the public
consciousness with remarkable speed and intensity.
Its current and potential applications affect virtually all
domains, from life-saving medical discoveries to smart
agriculture. AI is challenging the ways we think about
the world, and international trade is no exception,
as AI promises to transform trade logistics and supply
chain management and give rise to new forms
of services.

As this report shows, AI has the potential to reduce trade

costs, enhance productivity across sectors, and reshape
traditional trade patterns. I often say, the future of trade is
services; digital; and green; and that it must be inclusive.
AI can accelerate trade’s journey into this future.

The digital transformation driven by AI is poised not only data governance and privacy, how to regulate AI-enabled
to boost services trade; it may also create whole new products and associated ethical and societal risks, as
categories of tradable AI-powered goods, from autonomous well as how to protect intellectual property in an AI-driven
vehicles to robotics and beyond. If we successfully age. While we still need to find good answers to many of
harness its potential, AI can also support greener trade by these questions, it is already clear that making the most of
optimizing resource use and reducing the carbon footprint AI-related opportunities will require their benefits to be
of supply chains. widely shared across different economies.

But the inverse is also true. AI raises significant challenges, WTO economists simulated various AI uptake scenarios
from the growing risk of an “AI divide” to questions around for this report, and the differences were substantial.

4
Under an optimistic scenario they dub “global synergy”, Because of the remarkable pace at which AI is evolving,
in which AI is adopted evenly across regions and we need to look beyond today and anticipate what lies on
contributes to strong productivity gains, cumulative the horizon. This is why the report contains views from
real growth in global goods and services trade would scholars working at the intersection of AI, trade and the
increase by almost 14 percentage points through 2040, multilateral trading system. I want to emphasize that these
with global trade in digitally delivered services nearly 18 views and lines of inquiry do not reflect official positions
percentage points higher than the baseline projection. or carry the endorsement of WTO members or the
Conversely, under a cautious “tech divergence” scenario – Secretariat. They are in the report because they put before
characterized by divergences across regions in terms us some complex issues and difficult questions that we
of productivity increases and AI adoption – the impacts cannot afford to sidestep, and that should be read as an
of AI on trade growth would be halved, with a cumulative invitation for reflection and further research that will help
boost of only 7 percentage points by 2040. In other words, us better understand the fast-changing technological
failing to diffuse AI technology across different economies landscape in which the multilateral trading system
would mean foregoing many of the potential gains. operates. They may also serve as an inspiration for
discussion on the role of the WTO in supporting international
This report aims to stimulate a discussion on how the AI governance efforts.
WTO can promote the development and deployment of
AI and help mitigate its associated risks and looming By working together to leverage AI responsibly, we can
concerns about regulatory fragmentation. In this respect, drive sustainable economic growth, foster innovation,
two guiding questions the report tries to address are: and ensure that the benefits of this technology are
how can the WTO help ensure that the benefits of AI shared by all.
are broadly shared? How can the challenges that AI
presents be addressed in a globally coordinated manner? I invite all WTO members, stakeholders, and the broader
international community to engage with the findings of
The WTO matters here not just because of its rules and this report and to contribute to the ongoing discussions
adjudication functions, but also its role as a global forum on AI governance, including through the lens of trade
for discussion, coordination and cooperation. As the policy. Together, we can shape a future where trade and
report notes, this latter role is particularly relevant and technology work hand-in-hand to create a more prosperous,
suitable for AI: a complex and fast-evolving technology that sustainable and equitable world.
is inherently global in nature.

Of course, AI-related global governance cannot be

reduced to trade issues. But trade rules and policy
have important roles to play. The WTO, with its 166
members of all sizes and income levels, is well placed to
participate in ongoing AI global governance debates.
As the report notes, WTO members themselves are
slowly bringing AI into the agenda of various of our Dr Ngozi Okonjo-Iweala
deliberative bodies. Director-General

5
 Executive summary
The widespread and transformative impact that for developing economies and small businesses, helping
artificial intelligence (AI) is currently having on them to overcome trade barriers, enter global markets and
society is being felt in all areas, from work, production participate in international trade.
and trade to health, arts and leisure activities.
New applications of AI are expected to create unprecedented AI can transform patterns of trade in services,
new economic and societal opportunities and benefits. particularly digitally delivered services. It can enhance
However, significant ethical and societal risks are also productivity, especially in services sectors that rely on
associated with the development and application of AI. manual processes, by enabling low-skilled workers to
These risks have implications for all these areas too, leverage best practices of more high-skilled workers more
including trade. AI is a global issue, and as governments effectively. For example, generative AI can amplify the
increasingly move to regulate AI, global cooperation is more performance of business consultants by up to 40 per cent
important than ever. compared to those not using it. Greater productivity gain
is also observed for lower-skilled workers (Dell’Acqua et
Against this backdrop, the present report examines al., 2023). Research also shows that access to generative
the intersection of AI and international trade. AI increases the productivity of call centre workers by an
It begins with a discussion of why AI is a trade issue, average of 14 per cent, and by 34 per cent specifically for
before delving into the ways in which AI may shape the novice and low-skilled workers (Brynjolfsson et al., 2023).
future of international trade. It discusses key trade-related AI can also foster the development of innovative services
policy considerations raised by this technology and and increase demand for them. However, while AI can
provides an overview of government initiatives taken both enhance trade in digitally delivered services significantly,
to promote and to regulate AI. The report also highlights the it has contributed to reducing the demand for certain
looming risk of regulatory fragmentation and its impact, in traditional services. AI-enabled automation can also reduce
particular on trade opportunities for micro, small and medium- the necessity to outsource certain services.
sized businesses. Finally, the report discusses the critical
role of the WTO in facilitating AI-related trade, ensuring AI can increase demand and trade in technology-
trustworthy AI and addressing emerging trade tensions. related products. Because AI systems often rely on
real-time data streams and seamless connectivity, the
adoption of AI is spurring demand for complementary goods
Why is AI a trade issue? related to information and communications technology (ICT)
infrastructure and information technology (IT) equipment.
AI is distinct from other digital technologies in These include computer and telecommunications services,
several key ways, and it has the potential to affect specialized development tools and software libraries.
international trade significantly. It is a general-purpose For example, the global market for AI chips was valued at
technology, capable of adapting to a wide range of domains US$ 61.5 billion in 2023 and it has been projected that it
and tasks with unprecedented flexibility and efficiency. It could reach US$ 621 billion by 2032 (S&S Insider, 2024).
relies on large datasets to learn and improve its performance As many of these goods and services are often supplied
and accuracy. AI's functions and efficiency can evolve by a small number of economies, international trade serves
rapidly, leading to dynamic shifts in its capabilities and as a major channel to foster AI development worldwide.
autonomy. Finally, its inherent complexity and opacity, as well Further upstream in the value chain, trade in the extraction
as its potential failures and biases, raise significant concerns and processing of critical metals and minerals, as well as
related to matters such as how to understand the reasons trade in energy, are also likely to gain in importance.
for and basis of AI decisions and recommendations, or In addition, AI has substantially heightened the demand for
regarding ethics and broader societal implications. data, fundamentally reshaping the landscape of data usage
and trade.
AI can be leveraged to overcome trade costs
associated with trade logistics, supply chain By affecting productivity, and through shifts in
management and regulatory compliance. By production dynamics, AI may reshape economies'
enhancing trade logistics, overcoming language barriers, comparative advantages. AI is expected to enhance
and minimizing search and match costs, AI can make productivity across all economic sectors in both developed
trade more efficient. It can help to automate and streamline and developing economies, and to change the composition
customs clearance processes and border controls, navigate of inputs required for production, placing greater emphasis
complex trade regulations and compliance requirements, on capital investment, rather than on labour inputs. This
and predict risks. AI-based tools can be used in trade shift in production dynamics could reshape trade patterns.
finance, and can significantly enhance supply chain visibility Conversely, new sources of comparative advantage may
by providing real-time data analytics, predictive insights and emerge from factors like educated labour, digital connectivity
automated decision-making processes. All of this could and favourable regulations. Because AI is energy-intensive,
lower trade costs and, as a result, level the playing field economies with abundant renewable energy may also

6
gain comparative advantages. However, although AI can as are data governance challenges and the need to
potentially benefit all economies, the development and ensure that AI is trustworthy and to clarify how it relates to
control of AI technology are likely to remain concentrated intellectual property (IP) rights. The implementation of AI
in large economies and companies with advanced AI at the domestic, regional and international levels entails
capabilities, resulting in industrial concentration. both benefits and risks, and a lack of coordination could
cause increasing regulatory fragmentation with regard to AI.
The adoption of AI can drive productivity increases
across various sectors and reduce trade costs, Addressing the risk of a growing AI divide is essential
leading to global gains in trade and GDP. Simulations to leverage the opportunities offered by this
using the WTO Global Trade Model show that, under an technology. Currently, the capacity to develop AI
optimistic scenario of universal AI adoption and high technology is concentrated in a few large economies, and
productivity growth up until 2040, global real trade growth this is creating a significant divide between economies
could increase by almost 14 percentage points. In contrast, that are leading research and development (R&D) in AI –
a cautious scenario, with uneven AI adoption and low in particular China and the United States – and the rest
productivity growth, projects trade growth of just under of the world. This imbalance could be further exacerbated
7 percentage points. The simulation further shows that, by the use of government subsidies to develop AI. The risk
while high-income economies are expected to see the of industry concentration within a few large firms could
largest productivity gains, lower-income economies have also intensify the divide between firms. These features,
better potential to reduce trade costs. combined with the opacity of AI algorithms and the
possibility of tacit collusion among competitor firms
The global trade and GDP impact of AI varies to maintain higher prices, present challenges for
significantly across economies and sectors, competition authorities.
depending on choices made concerning innovation
and policies. While trade growth in high-income economies The rise of AI is raising important data governance
remains relatively stable across projected scenarios, issues that will need to be addressed to prevent
low-income economies could experience much higher trade further digital trade barriers. Cross-border data flows
growth under the scenarios of universal AI adoption and high are essential to AI, as vast amounts of data are needed to
productivity growth (18.1 percentage points) compared to train AI models, as well as minimize possible biases.
those of uneven AI adoption and low productivity growth Thus, restrictions on data flows can slow AI innovation
(6.5 percentage points). The simulation results suggest and development, increase costs for firms, and negatively
that if developing economies improve their AI readiness by impact trade in AI-enabled products. A recent study
strengthening digital infrastructure, enhancing skills and (OECD and WTO, 2024) found that if all economies
boosting innovation and regulatory capacity, they will be in a fully restricted their data flows, this could result in
better position to adopt AI effectively. a 5 per cent reduction in global GDP and a 10 per
cent decrease in exports. However, the large datasets
These simulations show that digitally delivered required by AI models raise significant privacy concerns.
services1 are expected to experience the highest Therefore, a reasonable trade-off between accessing
trade growth. In an optimistic scenario of universal AI large amounts of data to train AI models and protecting
adoption, digitally delivered services are projected to see individual privacy must be found.
cumulative growth of nearly 18 percentage points relative
to the baseline scenario, the largest increase across all Ensuring that AI is trustworthy without hindering
sectors. The expected impact of AI on real trade growth trade can be challenging. “AI trustworthiness” means
also differs within sectors. Potentially digitally delivered that it meets expectations in terms of reliability, security,
services such as education, human healthcare, and privacy, safety, accountability and quality in a verifiable
recreational and financial services, as well as manufacturing way. However, given the behaviour and opaque nature
sectors such as processed food, are projected to of AI systems, as well as the potential dual-use of some
experience significant trade growth, largely driven by trade AI products (i.e., for both civilian and military applications),
cost reductions. Meanwhile, sectors related to natural striking a balance between ensuring that AI is trustworthy
resource extraction and manufacturing sectors such as and enabling trade to flow as smoothly as possible may
textiles are expected to see limited growth. prove especially challenging. The evolutionary nature of
AI makes regulation a perennial moving target. “Traditional"
regulations and standards for goods, which normally focus
The policies of AI and trade on tangible, visible and static product requirements, may
not be fully capable of addressing all of the different types
The discussion on how AI might reshape international of potential risks, including the ethical and societal
trade raises important policy questions. The risk of a questions that may result from the integration of AI into
growing divide resulting from applications of AI is significant, goods and services. Regulating to address questions

7
 of public morals, human dignity and other fundamental For example, while some bilateral cooperation initiatives
rights, such as discrimination or fairness, is not only focus primarily on aligning AI-related terminology and
challenging, but is also prone to causing regulatory taxonomy, and on monitoring and measuring AI risks,
fragmentation because the meaning and relative importance others prioritize collaboration to promote alignment in
of such values may vary across societies. general terms or focus primarily on AI safety and governance.
Likewise, some regional initiatives prioritize human rights
AI also poses new conceptual challenges for and ethics, while others focus on economic development
the traditional, “human-centric” approach to IP rights. and growth.
Issues that deserve particular attention include the
protection of AI algorithms and of copyrighted material Regional trade agreements (RTAs) and digital
for training AI, and the protection and ownership of economy agreements are important vehicles to
AI generated outputs. These questions may call for a promote and regulate AI. AI-specific provisions have
re-evaluation of existing IP legal frameworks. started to be incorporated into such agreements, but they
mainly take the form of “soft” – i.e., non-binding – provisions
The immense potential of AI has prompted focusing on the importance of collaboration to promote
governments around the globe to take action to trusted, safe and responsible use of AI. Several AI-specific
promote its development and use while mitigating provisions explicitly refer to trade. Digital trade provisions
its potential risks. At the domestic level, more and more included in RTAs, such as provisions on data flows, data
jurisdictions are putting in place AI strategies and policies localization, protection of personal information, access
to enhance their AI capabilities. The number of economies to government data, source code,2 competition in digital
having implemented AI strategies increased from three in markets, and customs duties on electronic transmissions,
2017 to 75 in 2023. According to Stanford University's are also important for AI development and use. The number
2024 "AI Index", 25 AI-related regulatory measures were of RTAs with digital trade provisions has been growing
adopted in the United States in 2023, compared to just steadily since the early 2000s, and by the end of 2022,
one in 2016, while the European Union has passed almost 116 RTAs – representing 33 per cent of all existing RTAs
130 AI-related regulatory measures since 2017. However, – had incorporated provisions related to digital trade
most domestic AI policy initiatives are being implemented (López-González et al., 2023). However, the depth of digital
by developed economies, which could further deepen trade provisions included in RTAs varies significantly,
the existing AI divide between developed and developing reflecting diverging approaches. Few developing economies
economies: while around 30 per cent of developing and LDCs have negotiated digital trade provisions.
economies have put AI policy measures in place, only one Disciplines on trade in services in RTAs are also an important
least-developed country (LDC) – Uganda – has done so channel through which governments' trade policies and
according to data from the Organisation for Economic trade obligations can affect the policy environment for
Co-operation and Development (OECD) AI Policy AI, but the level of commitments undertaken differs
Observatory. Also high on governments’ policy agendas are significantly across economies.
domestic initiatives to promote access to data through
open data and data-sharing initiatives, with a view to The last few years have witnessed a wave of
fostering domestic innovation and competition, protecting international initiatives related to AI. While there
privacy and controlling the flow of data across borders. are elements of complementarity among such initiatives
and alignment on core principles, different initiatives
What is emerging is a landscape of fragmented prioritize different aspects of AI governance. A number of
measures and heterogeneous domestic initiatives, initiatives also contain various common elements that have
which may lead to regulatory fragmentation. important trade and WTO angles, such as the recognition
This fragmentation extends beyond AI-specific regulations of the role of regulations and standards, the need to avoid
to include sector-specific legislation, such as IP and data regulatory fragmentation, the importance of IP rights, the
regulations, which also impact AI. In addition, the design importance of privacy, personal data protection and data
of some border measures imposed on the hardware governance, and the importance of international cooperation,
components and raw materials crucial to AI systems can coordination and dialogue. Several of these initiatives also
affect competitors in other economies, leading to trade- address the environmental impacts of AI.
distorting effects and further exacerbating fragmentation.
The economic costs of regulatory fragmentation, in However, there is still no global alignment on AI
particular for small businesses, highlight the importance of terminology. Differing priorities, the overlap between
mitigating regulatory heterogeneity; according to OECD initiatives, and lack of global agreement on key terminology
and WTO (2024), the economic costs of the fragmentation could pose challenges at the implementation stage,
of data flow regimes along geo-economic blocks amount limiting efforts to prevent fragmentation and to put in
to a loss of more than 1 per cent of real GDP. place a coherent global AI governance framework.
Nevertheless, beyond initiatives to govern AI, an increasing
The increasing number of bilateral and regional number of international organizations, such as the
cooperation initiatives on AI governance, many International Telecommunication Union (ITU), the United
focusing on different priorities, add to the risk of Nations Educational, Scientific and Cultural Organization
creating a multitude of fragmented approaches. (UNESCO), the United Nations Industrial Development

8
Organization (UNIDO) and the World Bank, are developing rulebook can contribute to promoting the development
courses on AI and integrating AI in their technical of and access to AI. For example, the General Agreement
assistance activities, some of which have a trade component. on Trade in Services (GATS) plays an important role in
shaping a policy environment that facilitates the
The WTO, as the only rules-based global body development and uptake of AI. A majority of WTO
dealing with trade policy, can contribute to promoting members (out of 141 schedules of commitments, 84,
the benefits of AI and limiting its potential risks. or 60 per cent, contain commitments on computer services)
It can play an important role in limiting regulatory have made specific commitments on market access and
fragmentation, promoting the development of trustworthy national treatment related to ICT services, which play a
AI and access to it, and facilitating trade in AI-related goods fundamental role in enabling and promoting AI. However,
and services, thereby enabling the growth of AI and commitments in other sectors remain limited, and barriers
promoting innovation through IP. to services trade remain high in overall terms. When it
comes to goods, the Information Technology Agreement
(ITA) aims to increase worldwide access to high-
What role for the WTO? technology goods essential to AI by eliminating tariffs on
the ICT products it covers. Meanwhile, the TBT Agreement
WTO rules and processes promote global can help to ensure that, when governments adopt
convergence. The WTO is a forum that promotes AI standards and regulations, these are, to the extent
transparency, non-discrimination, discussion, the exchange possible, not trade-restrictive, and are optimal for
of good practices, regulatory harmonization, non-mandatory attaining policy objectives. The Trade-Related Aspects of
policy guidance, and global alignment through the Intellectual Property Rights (TRIPS) Agreement aims to
negotiation of new binding trade rules on trade. foster a balanced IP system that incentivizes innovation
Transparency provisions included in WTO agreements through the enforcement and protection of IP rights, while
allow WTO members, as well as economic operators promoting dissemination of and access to technology,
and consumers, to be kept abreast of latest regulatory to the mutual benefit of both producers and users of
developments. One example is the enhanced transparency technological knowledge. Various WTO agreements
provisions in the Technical Barriers to Trade (TBT) also include provisions to promote the transfer of
Agreement. By requiring early notification of regulatory technology, and this can play an important role in the
measures and allowing opportunities to provide development of AI. Finally, the WTO Agreement on
comments on these measures at a draft stage, the Government Procurement (GPA) 2012 promotes access
TBT Agreement can help to prevent obstacles to trade, to internationally available new AI technologies.
as well as promote and accelerate global convergence.
WTO members are increasingly notifying a wide range of Various principles, provisions and guidelines in
regulations on digital technologies to the TBT Committee. the WTO rulebook can support trade in AI systems
For instance, more than 160 notifications have been made and AI-enabled products by minimizing
on regulations addressing cybersecurity and the Internet international negative spillovers. Examples include
of Things (IoT)/robotics, both of which are relevant for AI. the non-discrimination principle and the Agreement on
More recently, the TBT Committee has started receiving Trade-Related Investment Measures (TRIMS), which
notifications of AI-specific regulations. Another example recognizes that certain investment measures can restrict
is the WTO Trade Policy Review Mechanism, which and distort trade and states that members may not apply
contributes to transparency in members’ trade policies. investment measures that discriminate against foreign
Finally, in terms of possible new substantive rules, various products or lead to quantitative restrictions. When it
issues negotiated under the Joint Statement Initiative on comes to technical regulations, standards and certification
E-commerce, which currently brings together 91 WTO procedures, the TBT Agreement provides that regulatory
members, may matter for AI. intervention shall not be discriminatory nor any more
trade-restrictive than necessary to achieve the intended
The WTO also provides a global forum for constructive policy objectives, and that it should, when justified, be
dialogue, the exchange of good practices, and subject to periodic reviews. And the Agreement on
cooperation. This enables discussion among members Subsidies and Countervailing Measures (SCM) can play
of how best to design nuanced, flexible and adaptable a crucial role in navigating the dual aspects of AI development,
regulatory solutions to address the goods, services and by promoting technological innovation while preventing
IP-related aspects of AI in a coordinated manner. In some negative spillovers in international trade from government
areas, the WTO also promotes regulatory harmonization financial support.
and coherence by encouraging the use of international
standards, mutual recognition and equivalence, and The WTO can help to prevent and settle trade
through various "soft law" instruments, such as voluntary tensions and frictions. The practice of raising "specific
committee guidelines.3 trade concerns" (STCs) allows WTO committees to
serve as a venue for defusing potential trade tensions
The WTO is the cornerstone of global efforts to with regulatory measures in a cooperative, pragmatic
facilitate trade in services and goods that enable and non-litigious way. In the TBT Committee, for
or are enabled by AI. Various aspects of the WTO instance, members have already been using this practice

9
 to discuss and address concerns with regulations further to help developing economies seize the benefits of
involving a wide range of digital technologies and issues, AI for trade.
including IoT, autonomous vehicles, 5G in robotics,
industrial automation, cybersecurity, and more recently As a forum for negotiation, discussion and
AI. The WTO also serves as a global forum to settle rule-making, the WTO provides a multilateral
trade-related disputes. While there has been no dispute framework that can help address the trade-related
on AI so far, the WTO Dispute Settlement System has aspects of AI governance. Nevertheless, AI may have
dealt with resolving disputes related to various aspects implications for international trade rules. Although it is
of the digital economy. a new technology, AI is developing rapidly, and is certainly
already advanced enough to be a subject of discussions
The WTO promotes inclusiveness through special at the WTO. Its cross-cutting nature requires a
and differential treatment and technical assistance cross-cutting policymaking approach to promote
for developing economies. WTO agreements recognize policy coherence.
the constraints faced by developing economies and, for
this reason, include various special and differential (S&D) While AI governance extends beyond trade, trade
treatment provisions to help them to implement WTO remains a crucial element within AI governance.
rules and participate more effectively in international trade. The WTO can contribute significantly to developing
Technical assistance and capacity-building are key pillars a robust AI governance framework. This report is a
of the WTO’s work and play a fundamental role in furthering first attempt to explore some key implications of AI for
understanding of the WTO rules and agreements, as trade and trade rules. As AI continues to evolve,
well as of other topics relevant to trade. Multi-stakeholder governments should continue to discuss the intersection
programmes, such as Aid for Trade and the Enhanced of AI and trade and its possible implications for the
Integrated Framework, could, however, be leveraged WTO rulebook.

Endnotes
1 Simulations in this report define digitally delivered services as 3 Such “soft law” instruments also include the set of Principles for the
services that can be delivered remotely over computer networks, Development of International Standards, Guides and Recommendations
(WTO et al., 2023). agreed by the TBT Committee in 2000 (the "Six Principles") and the TBT 2024
2 See Annex 1 for further explanation of key concepts in AI. Conformity Assessment Procedures (CAP) Guidelines.

10
1 Introduction
11
CHAPTER 1: INTRODUCTION

Introduction
With the launch of ChatGPT in November 2022, artificial understanding of the intersection between AI and trade in
intelligence (AI), and in particular generative AI – capable order to ensure that AI’s benefits for trade and economic
of generating high-quality text, images and other content growth are harnessed, and that related risks are mitigated.
based on the data on which it is trained – entered into public
consciousness and has been experiencing rapid adoption. This report discusses how AI impacts trade and how trade
and trade policies impact AI. It explores how AI may shape
AI is a general-purpose technology that is already the future of international trade and examines some of
having, and will continue to have, a pervasive impact on the key trade-related policy considerations that this
our societies. It encompasses a broad spectrum of technology raises.
technologies with numerous applications that have the
potential to transform deeply the way we work, produce It discusses how governments are responding to the
and trade. new opportunities and challenges raised by AI, and the
consequent potential risk of policy fragmentation, and it
Rapid advances in AI are expected to reduce trade costs, explores the role that the WTO can play in facilitating trade
boost productivity and innovation, and reshape economies’ in goods and services related to AI, promoting trustworthy
comparative advantages, creating unprecedented new AI and addressing trade tensions. Finally, it discusses
economic and societal opportunities and benefits. An possible implications of AI for international trade rules.
international trade environment prepared to facilitate these
changes is key to further developing AI and to reaping its
related benefits and opportunities.

Through the rules and commitments contained in its

agreements, the WTO can play an important role in this “Rapid advances in
context, by ensuring that goods using or supporting AI AI are expected to
can flow without encountering trade obstacles, and by
providing a conducive climate for trade and investment in reduce trade costs,
AI-related services.
boost productivity
However, AI is also giving rise to significant risks and and innovation, and
challenges. The fact that it can already be used and applied
globally means that any action taken to develop, apply and
reshape economies’
control it must also take place at a global level. comparative
As the broader regulatory and policy landscape surrounding advantages.”
AI is taking shape, it is critical to establish a better

12
2 Why is AI a
trade issue?
13
CHAPTER 2: WHY IS AI A TRADE ISSUE?

content, recommendations, or decisions that can influence

(a) What is artificial physical or virtual environments. Different AI systems vary in

their levels of autonomy and adaptiveness after deployment”
intelligence (AI) and (OECD, 2024a). As for the International Organization for

what makes it unique? Standardization (ISO), it defines an AI system as an

“engineered system that generates outputs such as content,
forecasts, recommendations or decisions for a given set of
human-defined objectives”.1
AI encompasses a broad spectrum of technologies
with numerous applications. There are several definitions While the history of AI as a research field began in
of AI systems by international bodies. The Organisation for the 1950s, recent decades have seen unparalleled
Economic Co-operation and Development (OECD), for growth in AI applications. As illustrated in Figure 2.1, the
example, defines an “AI system” as “a machine based system field was initially met with enthusiasm, leading to the creation
that, for explicit or implicit objectives, infers, from the input of programmes that could play chess and solve algebraic
it receives, how to generate outputs such as predictions, problems. However, progress slowed during an “AI winter”

Figure 2.1: A brief history of AI

1950:
Visionary computer scientist Alan Turing
suggests a language-based test to evaluate
whether a machine has the ability to exhibit
intelligent behaviour equivalent to,
or indistinguishable from, that of a human
being: the Turing Test is invented. 1956:
The term “artificial intelligence” is coined
during a seminal workshop at Dartmouth
College, United States.
1950s-60s:
Work focuses on the use of logic by
symbolic AI – which processes symbols
or concepts, rather than numerical data – 1970s-80s:
to imitate human intelligence. Expert systems, which emulate the
decision-making abilities of human experts,
have a period of popularity, followed by the
“AI Winter”, resulting from limitations in
1997: computing power and problem complexity.
IBM’s “Deep Blue”, a chess-playing
computer system, defeats chess champion
Garry Kasparov, showcasing AI’s potential
for complex decision-making. 2012:
Breakthroughs in deep learning advance
computer vision, natural language
processing and speech recognition.
2017:
Google’s AlphaGo defeats Ke Jie, the world
champion of the board game Go, demonstrating
the potential of deep learning. 2010s-present:
AI becomes broadly available through
open-source tools and cloud computing.

2022:
The public launch of Chat GPT3 brings
generative AI to the attention of the 2024:
general public.
Development of AI ethics and regulatory
frameworks to ensure its responsible
application and use.

14
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

period of reduced funding and interest. Renewed It is increasingly playing a role in every sector of the economy
advancements in the 1980s, followed by breakthroughs in and in every aspect of our daily lives. From driving our cars
machine learning (i.e., the ability of machines to learn without to controlling our critical infrastructure, diagnosing our
explicit programming) and neural networks (i.e., a type of illnesses and recommending content for our entertainment,
machine learning by which a computer learns to perform AI is ubiquitous (Shadbolt, 2022), leading some to term AI
a task by analysing examples) in the 2000s, have since driven an “omni-use” technology (Suleyman and Bhaskar, 2023).
AI to its current prominence and its increasing application AI technologies are prevalent across various domains, such
in various industries and in many people’s daily lives (see as language processing, vision (e.g., image recognition),
Annex 1 for further explanation of key concepts in AI). and multimodal systems that integrate and interpret more
than one type of data input. The number of AI systems
Contemporary advances in generative AI render AI applied in these domains has grown substantially in recent
distinct from other technologies in several key ways. decades (see Figure 2.2). As detailed in Box 2.1,
First, AI serves as a general-purpose technology, capable of AI can contribute to addressing environmental challenges
adapting to various domains and tasks with unprecedented and promoting sustainability.
flexibility and efficiency. Second, it feeds on large datasets
to improve its performance and accuracy. Third, its functions The fact that AI can be applied broadly means
and efficiency can evolve rapidly, leading to dynamic shifts that it can potentially be implemented both for
in its capabilities and applications. Finally, AI’s inherent beneficial and for harmful purposes. As a general-
complexity and opacity raise significant concerns regarding purpose technology, AI is particularly prone to misuses
ethics and broader societal implications. and dual uses (i.e., for both civilian and military
applications). For example, AI algorithms initially designed to
enhance productivity and optimize resource allocation
(i) A
 I is a general-purpose can also be repurposed for malicious ends, such as
illicit surveillance or misinformation campaigns. AI
technology with systems or models initially intended for civil use can
wide current and be repurposed for military uses, such as the
development of autonomous weapons systems.2 To many,
potential applications this underscores the critical importance of responsible
innovation, ethical AI governance and the establishment of
robust regulatory frameworks, to ensure that AI technologies
AI exhibits versatility in its capabilities, as it can are developed and deployed in ways that prioritize the
be applied to a wide range of tasks and domains. common good.

Figure 2.2: Domains of notable AI systems

200

180

160

140
Number of AI systems

120

100

80

60

40

20

0
1950
1952
1954
1955
1956
1957
1959
1960
1961
1962
1968
1970
1974
1975
1976
1977
1979
1980
1981
1982
1983
1984
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023

Language Vision Multimodal Audio Video Robotics

Recommendation Image generation Biology Speech Games Other

Source: Our World in Data based on Epoch (2024), last updated October 2024

15
CHAPTER 2: WHY IS AI A TRADE ISSUE?

Box 2.1:
The environmental impacts of AI

As a general-purpose technology, Moreover, it can help to measure, For instance, the training of
AI has the potential to help achieve simulate and reduce the environmental ChatGPT2, an earlier version of
a wide range of global sustainability footprint of supply chains (see Box 2.4 OpenAI’s language model released
goals. But AI also raises concerns and Barteková and Börkey (2022)). in February 2019, was estimated
regarding its potential adverse effects to produce 300 metric tons of
on the environment. Certain AI models can play an CO2 emissions, the equivalent
important role in addressing climate of 125 round trip flights between
The potential environmental benefits adaptation and resilience. They are New York and Beijing (Strubell et
of AI are manifold. For instance, it can increasingly capable of weather al., 2019). The computational and
reduce the energy carbon footprint forecasting and enhancing severe environmental costs of training can
by improving the efficiency of smart event prediction, including tracking grow in proportion to the size of
electricity grids, complex supply tropical cyclones, atmospheric rivers the model (European Commission
chains and transport operations. In (i.e., moisture-carrying sections of et al., 2021). Furthermore, during
particular, when coupled with other the Earth’s atmosphere) and extreme their operational cycle, AI systems
emerging technologies, such as temperatures (Lam et al., 2023; can consume significant volumes
synthetic biology (i.e., the design, Stanford University, 2023). In addition, of water, either directly, for cooling
engineering and modification of AI can enhance the efficiency and towers, or indirectly, through water
biological systems), and advanced reliability of renewable energy systems use for electricity generation. Some
materials, such as those used in by better understanding the supply predict that by 2027, the total water
nanotechnology,3 AI can foster a whole and demand dynamics, maximizing the consumption of all AI systems may
new wave of revolutionary innovations financial value of renewable energy exceed 0.38–0.60 billion cubic
(Stanford University, 2023).4 and allowing it to be integrated more metres, roughly 200,000 Olympic-
easily into the grid (IEA, 2023). sized swimming pools (Ren, 2023).
AI could also improve greenhouse A study indicates that data centres,
gas absorption and carbon storage by However, AI can also result in both cryptocurrencies and AI consumed
monitoring and predicting emissions direct and indirect negative impacts almost 2 per cent of total global
from ecosystems (OECD, 2022). It on the environment. Direct impacts electricity demand in 2022 and
can facilitate sustainable trade and stem from the use of resources these figures could double by
protect biodiversity by means of tools throughout the AI system’s lifecycle. 2026 (IEA, 2024).
such as image-based detection of Particularly impactful is the
illegal wildlife trade, high-risk animal consumption of resources such as The environmental impacts of
tracking, food value chain optimization water, energy and other raw materials, AI are being addressed in
and source monitoring and tracking and the associated greenhouse gas several government and
(World Economic Forum, 2018). emissions (OECD, 2022). intergovernmental initiatives.

(ii) AI feeds on large In sum, data provide the raw material and fuel enabling AI
systems to train, learn and improve.
datasets and data
regulations play a The data utilized in AI applications can vary widely in
terms of its sourcing and accessibility. Some datasets
pivotal role in this are open-source and may be contributed by organizations,
researchers or individuals with the intention of fostering
innovation and collaboration within the AI community. On the
AI algorithms require vast amounts of data to learn other hand, proprietary data is owned and controlled by
patterns, make predictions and perform tasks specific entities, and access to these data may be restricted
accurately. The quality and quantity of data directly impact the and require agreements or licences for use. Proprietary data
performance and reliability of AI systems: high-quality, diverse sources can include internal company data, research
datasets enable AI models to generalize and adapt to new datasets or commercially acquired data. With the exponential
scenarios, supporting continuous iteration and improvement. growth in the volume and variety of data available to AI
Access to up-to-date, representative datasets is therefore systems, privacy and intellectual property (IP) concerns loom
crucial to keep AI systems relevant and effective over time. larger than ever (see Chapter 3(a)).

16
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

Data regulations play a pivotal role in determining often insufficient to predict the duration or trajectory of future
the use of data and shaping the process of AI use advancements. Therefore, policymakers aiming to stay abreast
and innovation. Many regulations establish guidelines for of advancements in AI technologies cannot solely rely on past
obtaining consent, providing transparency and safeguarding developments; they must adapt to and anticipate changes as
sensitive information (See Chapter 3(b)). Data regulations they arise.
take into consideration the balance between AI innovation
and deployment on the one hand, and the need for privacy
protection, ethical considerations, IP rights and data security (iv) A
 I’s inherent complexity
on the other hand. In an increasingly interconnected world,
data regulations also govern the cross-border transfer and
and opacity and its
sharing of data between jurisdictions. potential failures and
biases create challenges
(iii) AI’s functions can for regulators
evolve rapidly,
leading to dynamic AI models often exhibit a significant degree of opacity

shifts in its capabilities in their decision-making processes. Deep learning models

(see Annex 1 for explanation of key concepts in AI), in particular,
and autonomy operate through layers of algorithms and vast datasets,
resulting in a “black box” phenomenon where the rationale
behind specific outputs remains unclear to users and even
A key element of AI systems is their ability to to the designers of the deep learning models (Castelvecchi,
make significant, continual improvements to their 2016). Recent advancements in AI tools have empowered
performance. This is attributable to, and dependent on, three machines to tackle tasks that go beyond explicit, fully specified
primary factors: algorithmic innovation, data availability and sets of rules and procedures, further exacerbating concerns
computational resources. Algorithmic advancements have about their lack of transparency.
paved the way for more sophisticated and effective AI models.
The abundance of high-quality data is providing AI systems with The opacity of AI models can lead to challenges in
rich and diverse information to learn from, while the exponential understanding how they arrive at their decisions or
growth in computational power is empowering researchers and predictions. While researchers and practitioners are actively
practitioners to train larger and more complex models at scale. exploring various techniques to shed light on the decision-
making processes of AI systems, through initiatives such
As AI models become more sophisticated and datasets as “Explainable Artificial Intelligence (XAI)”,6 this lack of
grow larger, the quantity of computing resources used transparency raises ethical and accountability concerns. There
in AI training increases exponentially. It is estimated that is a lack of standardization in how AI models are developed,
the computational resources needed to train AI have doubled documented and evaluated. This variability across models and
every 3.4 months since 2012 (Amodei and Hernandez, 2024). applications further complicates efforts to make AI systems
This progress is driven by the willingness of industries to use transparent and understandable (Ananny and Crawford, 2018).
more data centre capacity for large-scale general-purpose AI
training. This can be compared to the processing power of Risks of malfunction, misinformation and bias in AI
computer chips which, since the 1960s, has tended to double could have significant ethical and societal impacts. AI
approximately every 18 to 24 months, a phenomenon famously algorithms used in decision-making processes can perpetuate
known as “Moore’s Law” after Gordon Moore, one of the biases present in historical data, leading to unfair outcomes and
cofounders of Intel Corporation.5 reinforcing systemic inequalities. In trade, biased AI systems
can unfairly disadvantage certain groups or economies; for
AI can exhibit varying levels of autonomy, depending example, misclassifying businesses from specific regions
on its design and purpose. AI systems can range from as high-risk can limit market access for these businesses.
supervised systems that require human oversight to fully In addition, while AI can optimize global supply chains, it
autonomous systems capable of independent decision-making. may prioritize cost savings over ethical practices, leading to
Some AI systems learn from data and adjust their behaviour reputational risks and potential sanctions.
based on experience, while others operate independently in
real time without human intervention, particularly in domains like Addressing these challenges requires concerted
autonomous vehicles and robotic automation. As AI systems’ efforts. It involves developing robust mechanisms for
autonomous capabilities evolve and increase, concerns with detecting and countering the spread of false information,
the need to ensure human agency and oversight grow. ensuring transparency and accountability in AI algorithms,
and promoting diversity and inclusivity in dataset collection
The need to keep pace with the swift evolution of AI is and model development. In addition, it is crucial to carry out
creating challenges for regulators. AI’s rapid advancement continual evaluations of the quality of training data for AI
suggests that its capabilities may grow exponentially in the systems, including the adequacy of the data collection and
future. While past trends offer valuable insights, they are selection processes, proper data security and protection

17
CHAPTER 2: WHY IS AI A TRADE ISSUE?

measures, and feedback mechanisms to learn from mistakes AI systems have the capacity to enhance international
and share best practices among all AI actors (UNESCO, communication. Through AI-driven virtual collaboration
2021). Fostering digital literacy and critical thinking skills tools, including advanced video conferencing with features
among users can help to mitigate the impact of misinformation like noise cancellation and automatic transcription and
and bias in AI-driven technologies. translation, as well as virtual and augmented reality, seamless
communication and collaboration among global teams
and partners are facilitated. AI can significantly enhance the
functionalities of information and communications technology
(b) H
 ow will AI affect (ICT) services, which can enable businesses to overcome
geographical barriers and engage in real-time interactions,
international trade? negotiations and decision-making processes, facilitating
international trade and reducing the need for physical travel.

This section discusses how AI may reshape the future AI can significantly reduce search and match costs
of trade. It addresses questions as to how AI may be in trade by streamlining the process of identifying
used to overcome trade costs, how it can alter the pattern potential trading partners. AI-powered search algorithms
of trade in services, how it can affect trade in certain goods, can efficiently sift through vast amounts of data from
and how it may affect economies’ comparative advantages. various sources to identify potential trading partners,
suppliers, buyers and distribution channels. AI-driven
recommendation systems can analyse historical transaction
(i) AI holds the potential data, user preferences and market trends to provide
to significantly reduce personalized recommendations for potential trade
opportunities. AI-powered marketplace platforms can
trade costs facilitate matchmaking, automate contract negotiations
and optimize pricing strategies based on supply and
demand dynamics.
AI can reduce trade costs
by enhancing trade logistics, AI serves a multitude of purposes
overcoming language barriers in customs and border controls
and minimizing search costs.
As huge volumes of data are generated by people
AI technologies are revolutionizing supply chain and goods moving across borders, AI can be used
management by optimizing inventory management, for a range of purposes in customs and border
demand forecasting and logistics. As illustrated in controls. This includes optimizing revenue collection
Figure 2.3, by collecting and analysing data from various models to ensure accurate tax and duty collection, simplifying
sources, including Internet of Things (IoT) devices, AI systems product classification under the Harmonized System (HS)8
can generate insights into historical data, market trends for enhanced compliance, facilitating faster anomaly
and external factors in order to predict demand, optimize identification during customs audits and enabling risk-based
inventory levels and improve order fulfilment. By using targeting of commercial shipments, for example using
AI to facilitate real-time tracking and monitoring of shipments, augmented/mixed-reality glasses for contraband detection
it is possible to provide better visibility, resulting in a (WCO-WTO, 2022).
reduction in delays and an increase in efficiency. For example,
AI enables commercial shipping companies to predict Many customs administrations are using or plan to
ship arrivals five days in the future with high accuracy, use AI. According to a survey by the World Customs
thus enabling real-time allocations of personnel and Organization (WCO) and the WTO, 25 per cent of
schedule adjustments.7 respondents currently utilize AI and machine learning in
customs administration, with an additional 25 per cent
AI systems can eliminate language barriers by intending to implement them. The primary reported benefits
providing real-time translations. Various AI-driven include improved risk management and profiling, enhanced
language translation systems, powered by deep learning fraud detection and compliance, and more effective
techniques, can provide real-time translation services, customs audits for identifying anomalies (WCO-WTO,
facilitating seamless communication between speakers of 2022). As discussed in detail in Box 2.2, AI can serve
different languages regardless of their native tongue. Having multiple functions in streamlining and improving the
this level of connectivity facilitates smoother negotiations accuracy of customs processes.
and collaborations and the sharing of vital information,
fostering stronger global ties. A study shows that the AI-powered automated detection tools can greatly
introduction of a new machine translation system in a digital facilitate the work of customs officials. The application
platform has resulted in a remarkable 10.9 per cent increase of AI to customs risk assessment enhances the security and
in international trade between pairs of economies where efficiency of border crossings, allowing for the identification
people used this new system (Brynjolfsson et al., 2019). of potential risks and anomalies in shipments, and

18
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

Figure 2.3: How AI can enhance supply chain management

Intelligent workflow Intelligent manufacturing

and processes • Can predict potential
• Enhanced data quality: can help structure disruptions and optimize
data, create connections, identify partners production
• Optimization of processes for greater • Automation through
operational effectiveness AI-enabled robots
• Can help identify trading partners and assess • Product anomaly detection
compliance (including compliance of suppliers)
• Eliminates language barriers and
enhance communication

Intelligent
tracking
Can track materials, products,
carbon emissions, especially
when combined with
blockchain and tracking
technologies

Intelligent customs Intelligent inventory

• Risk management, profiling, Can predict demand,
enhanced fraud detection optimize inventory,
and compliance predict potential disruptions
• Tariff classification
• Revenue collection Intelligent transportation
• Post-clearance audits and logistics
and control Can optimize routes, reducing
environmental impact, transportation
time and fuel consumption
Source: WTO

19
CHAPTER 2: WHY IS AI A TRADE ISSUE?

Box 2.2: Case study:

Harnessing AI for enhanced trade facilitation
and border control at Dubai Customs
Dubai Customs has launched several Dubai Customs also launched the Last but not least, Dubai Customs has
projects leveraging AI to enhance Robotic Process Automation Smart launched a remote inspection initiative
trade facilitation and border control. Refund System to automate the that allows companies with the status
For example, iDeclare allows travellers claim and refund processes. The of authorized economic operator – a
to submit their customs declarations system uses AI to perform repetitive status granted by customs authorities
electronically and securely. office tasks, such as extracting to companies meeting security and
Passengers wishing to pre-declare data, filling in forms and moving compliance standards, allowing them
goods can upload a photograph files, and to match and validate to benefit from expedited customs
of the items to the application. The transaction details with minimal processes – to ask Dubai Customs to
app then selects the appropriate human intervention, thereby improving conduct inspections of their premises
HS code and determines whether transparency and reducing costs. using AI-powered robots equipped
and which customs duties are due. with thermal and infrared cameras.
Complementing iDeclare is the Al Post-clearance audits is another area
Munasiq app, a tool that assists users where Dubai Customs is leveraging Looking ahead, Dubai Customs is
in identifying the correct HS codes for AI to enhance customs processes exploring a wide range of additional
their items. Once the user enters the by automating the audit procedures AI applications, including automated
item’s description or photo, the app for high-value import declarations. threat detection algorithms, predictive
provides a ranked list of possible HS Several “bots” have been trained to analytics, AI-enabled drones for
codes along with relevant information, automate repetitive processes, such surveillance and machine vision
such as the description of the item, as data matching, and to interpret systems for inspecting containers.
applicable customs duty rate and any data and identify patterns, leading to
related prohibitions and restrictions. significant cost savings. Source: Musabih (2023).

enhancing security and efficiency at borders. In Brazil, and effectiveness of government procedures. By
for example, an AI system known as SISAM (“Sistema facilitating information-gathering on regulation changes
de Seleção Aduaneira por Aprendizado de Máquina”, or and automating compliance procedures, AI technology
“Customs Selection System through Machine Learning”) can help customs officials to stay abreast of evolving
has been leveraging the vast customs database to analyse regulatory landscapes with greater ease and efficiency.
each newly registered import declaration in the country. It can augment currently deployed digital solutions and
This system aids customs officers in identifying potentially allow for deeper automation, leading to improved efficiency
fraudulent customs declarations, thereby mitigating the risk and effectiveness of government control measures. For
of errors and enhancing compliance (WCO-WTO 2022). legislators, AI has the potential to simplify public commenting
processing on regulations and to improve the quality and
AI also offers opportunities to streamline en route richness of these comments.9
processes for customs clearance. For seaborne containers,
automatic detection transforms customs inspection into a Regulatory agencies have increasingly been using AI
streamlined process, significantly increasing inspection rates to predict risks and improve import screening.
without disrupting travel or trade. The Port of Qingdao in China, For instance, the US Food and Drug Administration (FDA)
for example, has installed a modular high-energy inspection employs the Predictive Risk-based Evaluation for Dynamic
system that scans every container along the sky rail route that Import Compliance Targeting (PREDICT) system to enhance
transports containers. This not only results in significant time import screening and targeting. This system aims to
savings and comprehensive security vetting, but also reduces prevent the entry of adulterated, misbranded or otherwise
the cost of container dispatching (Chen, 2022). violative goods into the United States, while expediting
the entry of compliant products. Similarly, in the
European Union, AI developments are crucial for tracing
AI can assist in navigating trade illegal activities within the agri-food chain, particularly
regulations and enhancing supply through the application of natural language processing.
chain visibility By leveraging AI to extract text from unstructured
databases and documents, these technologies can effectively
AI can assist in navigating complex trade regulations convert vast amounts of disparate data into structured,
and compliance requirements, improving the efficiency actionable intelligence.10

20
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

Box 2.3: Case study:

Benefits and challenges to the use of AI
for express delivery carriers
Express delivery carriers have with external data about where a more resilient, faster, and more
been experimenting with AI to parcel is travelling, such as data precise and reliable supply chains.
improve compliance, with two main about weather conditions, they can
objectives. The first objective is to better predict shipment delivery or Conversely, the main challenges that
better detect and challenge the provide better real-time intelligence express delivery carriers are facing
undervaluation (with the aim of for merchants about fulfilment or in deploying AI include the need to
paying fewer or no duties or taxes) of returns. balance the most viable technologies
declared goods, the misdeclaration with minimal infrastructure changes
of shippers or receivers, who wish Another application is testing to ensure that solutions are
thereby to bypass screening by AI-powered robots that have the customized for their business model.
the authorities, and incomplete or ability to see, touch, analyse and For instance, robotic solutions for
inaccurate goods descriptions, as move quickly to load trucks and warehouses with uniform boxes do
well as shipments of counterfeit or trailers with stable, dense walls of not work in an express environment
pirated goods. The second objective randomized boxes. where there are variations in the
is to validate client applications to size, weight, shape and packaging
open an account and to ensure that The main benefits derived from the materials of boxes. Building the
these clients represent a trustworthy use of AI in these contexts include right ecosystem requires quality
individual or company. better compliance levels, which infrastructure, talent and
help to reduce time at borders regulatory environments.
Express delivery carriers are also and to build trust with customs
using AI to improve processes. and other authorities, as well as Source: Based on information
For example, by combining better data-driven insights across provided by the Global
enhanced tracking information the company, which help to build Express Association.

AI can also greatly reduce the cost of business in patterns and anomalies using vast amounts of data from
complying with trade regulations. Through advanced various points along the supply chain. This enables companies
algorithms and machine learning capabilities, AI systems can to monitor inventory levels, track shipments and foresee
sift through vast volumes of regulatory documents, interpret potential disruptions with greater accuracy and speed.
intricate legal language, provide translation services and Moreover, as AI-powered tools can integrate data from
highlight pertinent updates or amendments relevant to trade disparate sources, they can offer a unified view of the supply
activities. As illustrated in Box 2.3, express delivery carriers chain, which can help to optimize logistics, reduce costs and
have been using AI to improve regulatory compliance. By improve overall efficiency. As illustrated in Box 2.4, enhanced
leveraging AI technologies, these carriers can more efficiently visibility through AI not only facilitates better strategic planning
manage and adapt to changing conditions and the dynamic but also supports more responsive and agile supply chain
regulatory environment. management. This could facilitate the compliance capabilities
of micro, small and medium-sized enterprises (MSME) to meet
AI-based tools can also be used in trade finance, and international trade regulations.
they are particularly useful for credit assessment,
risk evaluation and fraud detection. A multitude of data
sources are analysed in AI models to identify the Developing economies and small
creditworthiness of a business and provide a more accurate businesses benefit more from
risk profile by analysing financial records, market information AI-enabled trade cost reductions
and trade history. AI algorithms can also identify abnormalities
and patterns that indicate fraudulent activities, thus assisting Lower trade costs enable developing economies to
financial institutions in effectively mitigating the risks associated access global markets and participate in international
with those activities. trade. Historically, high trade costs, including tariffs,
transportation expenses and administrative burdens, have
AI can significantly enhance supply chain visibility by created significant barriers for developing economies seeking
providing real-time data analytics, predictive insights to export goods and services. However, AI and other digital
and automated decision-making processes. Through technologies can help to streamline trade processes and
advanced algorithms and machine learning, AI can identify diminish these barriers.

21
CHAPTER 2: WHY IS AI A TRADE ISSUE?

Box 2.4: Case study:

Using AI to improve supply chain
visibility and traceability
Multinational companies often In one instance, a pharmaceutical Companies can use AI-powered
struggle with significant blind company used the AI-enabled data engineering to automatically
spots in their product value value chain management system track emissions throughout their
chains, as they may be unable to identify the fact that a supplier carbon footprint and collect
to see beyond their direct who provides basic components data from operations, from
suppliers. This lack of visibility to suppliers higher up the supply activities such as corporate travel
can jeopardize the delivery chain had filed for insolvency and information technology (IT)
and reliability of their highest and ceased production after equipment, as well as from every
revenue-earning products. unsuccessfully restructuring its part of the value chain, including
business. The company was materials and components
An AI-enabled value chain able to mitigate the risk quickly suppliers, transporters, and
management system can address by alerting upstream suppliers even downstream users of their
this issue by providing and sourcing replacement products. AI can exploit data from
comprehensive insights into materials, thereby preventing new sources such as satellites,
all production steps, from raw a product shortage. In another can generate approximations of
material extraction to final goods case, a company faced a supply missing data and can estimate
distribution. AI connects and shortfall of a key plastic input for the level of certainty of the results.
learns from billions of data points, vaccine tubing. The AI-enabled Predictive AI can forecast future
offering detailed insights into facility value chain management system emissions across a company’s
geolocations, vendor profiles, found that the company’s direct carbon footprint. AI and
corporate ownership networks, supplier was a distributor, not optimization can improve efficiency
product transformations and a manufacturer, and that the in production and transportation,
third-party risk analytics, including disruption stemmed from an as well as in other areas, thereby
shipment dates, quantities, industry-wide plastic shortage. reducing carbon emissions and
geolocations and values. This insight helped the company cutting costs.
address the root cause and
Interactive maps constructed by seek alternative solutions. In summary, AI-driven systems
AI reflect the complexity of global can transform supply chain
supply chains, enabling proactive AI can also enhance a company’s visibility and traceability,
and reactive risk management. ability to gain insights into its carbon enabling companies to mitigate
For example, in vaccine footprint, promoting compliance risks, respond proactively to
manufacture, AI is used to provide with environmental, social and disruptions and achieve
real-time identification of supply governance (ESG) requirements sustainability goals, demonstrating
chain risks, such as exposure to and sustainable transformation. the profound impact of AI on
current events and bottlenecks According to a study by Boston value chain management and
that could introduce vulnerabilities Consulting Group (BCG), applying driving both economic and
in the future. This capability AI to corporate sustainability could environmental benefits.
allows for quick human generate up to US$ 2.6 trillion in
coordination and effective value through additional revenues Sources: Altana (2021) and
utilization of system outputs. and cost savings by 2030. Degot et al. (2021).

The reduction in trade costs levels the playing field to establish a global presence, form partnerships with overseas
for small businesses, helping them to overcome suppliers and distributors, and expand their customer base.
trade barriers and enter international markets. AI applications can automatically analyse, process and verify
Small businesses often face challenges like limited market data and provide integrated services for MSMEs, including
information, high transaction costs and complex trade automated processing with classification algorithms, error
regulations. AI-powered online marketplaces, digital marketing and fraud detection through anomaly detection, and capacity
strategies and e-payment systems enable small businesses planning using regression and forecasting (UNECE, 2021).

22
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

(ii) The most significant Recent research indicates that AI can substantially
enhance productivity, particularly for low-skilled
trade impact of workers, by leveraging best practices from other
AI will be on trade workers. With access to a large language model (LLM),
it is estimated that about 15 per cent of all worker tasks in
in services the United States could be completed significantly faster at
the same level of quality. When incorporating software and
tooling built on top of LLMs, this share increases to between
47 per cent and 56 per cent of all tasks (Eloundou et al.,
AI can boost productivity 2023). Within its operational scope, generative AI can amplify
in certain services sectors the performance business consultants by up to 40 per cent
compared to those not utilizing it (Dell’Acqua et al., 2023).
AI can enhance productivity, particularly in services A study of 5,000 workers responsible for complex customer
sectors that rely on manual processes. In these sectors, assistance at a call centre found that, among workers who
AI can significantly complement humans in improving were given the support of an AI assistant, the least skilled
efficiency, accuracy and the level of personalization (i.e., or newest workers showed the greatest productivity gain
the ability to tailor products, services, or experiences to (Brynjolfsson et al., 2023), while university-educated
meet individual preferences). Initial analysis suggests that professionals utilizing ChatGPT were more productive,
significant productivity gains are evident in sectors related efficient and satisfied with their tasks. Notably, individuals
to finance and insurance, management, information, and with weaker skills derived the greatest benefits from using
professional services (Figure 2.4). ChatGPT (Noy and Zhang, 2023).

Figure 2.4: AI can complement and enhance productivity in some occupations

Agriculture, forestry, fishing and hunting

Accommodation and food services

Mining, quarrying, and oil and gas extraction

Educational services

Transportation and warehousing

Construction

Manufacturing

Arts, entertainment and recreation

Health care and social assistance

Other services (except public administration)

Utilities
Administrative and support and waste
management and remediation services
Retail trade

Real estate and rental and leasing

Wholesale trade

Professional, scientific, and technical services

Information

Management of companies and enterprises

Finance and insurance

0.0 0.2 0.4 0.6 0.8 1.0

High AI impact (>50%) Low AI impact (10-50%) Minimal AI impact (<10%)

Source: Author’s elaboration based on an automation index developed by Eloundou et al. (2023).
Note: The figure shows the share of employment exposed to AI. High AI impact refers to sectors where the sector-level automation
index is 50 per cent or higher, low AI impact ranges between 10 per cent and 50 per cent, and minimal impact is less than 10 per cent.

23
CHAPTER 2: WHY IS AI A TRADE ISSUE?

AI can also foster the development Moreover, AI is shown to significantly enhance trade in
of innovative services, and boost digitally delivered services. By enabling the development
demand for them of more diverse mobile phone applications, AI has been
shown to increase the number of foreign users of AI-driven
AI’s capacity to derive valuable insights from extensive mobile applications by an average of tenfold (Sun and Trefler,
datasets is instrumental in fostering the development 2023). Similarly, the projections using the WTO Global Trade
of innovative services. In healthcare, for example, AI Model indicate that services in sectors such as education,
applications can significantly advance drug discovery and human health, recreation and finance could potentially
treatment methodologies, and may ultimately facilitate the undergo significant trade growth (see Section 2(b)(v)).
development of personalized healthcare solutions tailored
to individual patients. Similarly, AI-driven smart energy
management systems can integrate real-time sensor AI can automate and reduce the
data, weather forecasts, energy demand projections and demand for trade in certain services
equipment degradation profiles to provide dynamic
simulations, enabling energy companies to make informed, AI may contribute to reducing the demand for certain
proactive decisions. These systems optimize energy use, traditional services, as AI-driven automation can lead
reduce consumption and cut carbon emissions, resulting in to increased efficiency and productivity. For instance,
cost savings and improved sustainability. AI-powered legal research tools and contract review
systems can automate some tasks traditionally performed
In addition to fostering new discoveries, AI can also by legal professionals, potentially reducing the demand
enable customization of services to suit specific for certain legal services, especially in routine tasks like
preferences and use cases. By analysing vast amounts document analysis and discovery (OECD, 2024b). AI
of data to identify patterns and preferences, AI can allow chatbots and virtual agents have diminished the need for
for tailored solutions and adapt its outputs to meet the large customer service teams (see Box 2.5 on AI and jobs).
unique preferences of users. Examples include personalized
e-commerce recommendations, customized healthcare AI-enabled automation can reduce the necessity
treatments or individualized media content recommendations. to outsource certain services. According to recent
This customization not only enhances user satisfaction, surveys, companies have been using AI to streamline
but also enables the delivery of more targeted and effective manual or repetitive tasks and automate customer service
products and services across various industries. interactions (IBM, 2024). As a result, AI could reduce the
need for large call centres and business process outsourcing,
As AI becomes more integrated into daily life, services that many companies in developed economies
services that leverage AI capabilities to enhance often source overseas (Parkin and Kay, 2024). This could
convenience, efficiency and personalization are rising significantly impact developing economies, many of which
in demand. For instance, advancements in autonomous specialize in these types of services.
vehicles have paved the way for transportation services
such as ride-hailing platforms – matching passengers with
drivers for hire via online platforms – and delivery platforms,
which rely heavily on AI algorithms to optimize routes,
(iii) The emergence of
manage fleets and ensure safety. The rise of AI-powered AI will increase
virtual assistants, smart home devices and personalized
recommendation systems has fuelled demand for
demand and trade in
subscription-based streaming services. AI-powered AI-related products
recommendation systems in e-commerce platforms suggest
products based on users’ past purchases and browsing
history, driving increased sales and customer engagement. The adoption of AI technology is spurring demand
for complementary goods related to ICT
infrastructure and IT equipment. As illustrated in Figure
By bolstering productivity and 2.5, the AI value chain involves a range of products and
increasing demand, AI can boost services, and the rise of AI is likely to increase international
services trade trade in goods and services related to that value chain.

Increased productivity allows for greater output using AI applications, especially those involving deep
existing resources, thereby lowering production learning and neural networks, often require high-
costs. This phenomenon can spur heightened levels of trade performance computing systems to train complex
across diverse services sectors. Enhanced productivity models and perform intensive computations. Demand is
and innovation capacity can translate into increased trade rising sharply for hardware components of AI, such as
in certain services, leading to expanded trade volumes and high-performance CPUs (central processing units) and
enhanced economic interconnectedness on an international GPUs (graphics processing units) and specialized AI chips,
scale. As Richard Baldwin argues in his opinion piece, as well as switches and routers, which ensure fast data
AI could boost services trade in the future. transfers between systems. The global market for AI chips

24
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

Opinion piece Richard Baldwin

AI means that services will be the Professor of international
future of trade economics at the IMD
Business School
Global trade has long been dominated by
manufactured goods, but, as Bob Dylan sang back
in 1964, “The Times They Are a-Changin’.”

World exports of goods and services enjoyed

boomtime growth in the 1990s and early 2000s.
Since 2008, trade in goods – specifically manufactured First, barriers to intermediate services trade are
goods – have plateaued; services exports have not. technological, since there is almost no regulation
Services trade continues to ride the go-go growth of trade in back-office services. Second, digital
path it has been on since the 1990s. technology is lowering these barriers at an exponential
pace. Third, AI such as machine translation, and soon
Digital technologies in general, and AI in particular, simultaneous speech translation, are rapidly making
are why the times are a-changing, in my view. There domestic and foreign workers better substitutes than
are many reasons why manufactured trade slowed they were in the past. Generative AI (GenAI), I believe,
a decade and a half ago. This short essay skips over will accelerate this, since it levels up skills. GenAI
those reasons and jumps straight to how AI has distils the experience of a rich-nation services worker
spurred – and will continue to spur – services trade. into an app and then gives the app to emerging
economy services workers. The output of these
Digital technology, including communications, video low-wage workers will look a whole lot more like
conferencing and AI-driven machine translation, have that of G7 services workers when both G7 and
rapidly lowered barriers to trade in services. The emerging economy workers are using the same
changes that came with telework during the COVID-19 GenAI apps.
pandemic accelerated this trend by five to ten years.
Finally, the demand for intermediate services is
The main expansion has come in “intermediate huge in rich nations and the supply of appropriate
services”, which are the services sold by one business workers is huge in emerging economies, since
to another rather than to consumers. The ability to they are already providing these services in their
coordinate work teams across different locations local economies.
seamlessly has made it feasible for companies in
high-income economies to source services from What does all this mean? It is essential to recognize
emerging markets. For example, a US accountant that services – not goods – will be at the forefront of
might hire a bookkeeper in India to manage day-to-day global trade in coming years. The WTO Secretariat
accounting tasks. This arrangement is facilitated by needs to get ready since, “The Times They Are
digital tools that make remote collaboration easy, a-Changin’”.
cost-efficient and secure.
Disclaimer
Looking ahead, I see services trade growing faster Opinion pieces are the sole responsibility of their
than goods trade for the foreseeable future. I base authors. They do not necessarily reflect the opinions
this conjecture on four facts. or views of WTO members or the WTO Secretariat.

was valued at US$ 61.5 billion in 2023 and it has been are crucial to manage efficiently and access the vast
projected that it could reach US$ 621 billion by 2032 (S&S amounts of data required by AI systems. Fibre optic cables
Insider, 2024). are essential for high-speed data transmissions over long
distances. Sensors and actuators used in robotics and IoT
As AI systems often rely on real-time data streams applications are also in high demand.
and seamless connectivity, the demand for ICT and
network equipment will increase. Equipment such AI will also boost demand for computer and
as routers and switches is necessary to ensure high-speed telecommunications services, including software- and
internet connectivity and support AI-driven applications data-related services, as well as cross-border
and services. Hardware components like storage servers trade and investment in these sectors. Services to

25
CHAPTER 2: WHY IS AI A TRADE ISSUE?

Box 2.5:
How will AI impact jobs?

Unlike previous waves of impacts from AI. In economies compared to that generated by
technological transformation, with developed financial, legal capital investments. This shift
AI is poised to impact white-collar and technological sectors, the would benefit those who own
jobs more significantly than integration of AI into these capital and intellectual property
blue-collar ones. Historically, industries could lead to substantial (IP), or who have invested in
automation primarily affected changes in job dynamics. IMF AI-driven enterprises, and thus
manual labour and manufacturing research suggests that AI could it would further enrich already
jobs. However, AI’s capabilities endanger 33 per cent of jobs in wealthy segments of society.
extend into areas traditionally advanced economies, 24 per cent The concentration of wealth and
occupied by white-collar workers, in emerging economies, and 18 power in the hands of a few could
such as finance, legal services, per cent in low-income economies undermine democratic principles
and administrative roles (Cazzaniga et al., 2024). A and deepen existing power
(Autor, 2022). study by the International Labour imbalances within society.
Organization (ILO) predicts that
This shift means that roles the overwhelming effect of the In response to these challenges,
involving cognitive work, which technology will be to augment policymakers must proactively
were once considered more occupations, rather than to address the potential
secure from automation, are automate them, and the greatest consequences of AI on income
increasingly vulnerable to AI impact is likely to be in high and inequality. This may involve
technologies that can perform upper middle-income economies, implementing measures such as
these tasks faster and with greater due to a higher share of retraining programmes to equip
accuracy. Some economists employment in clerical occupations displaced workers with skills
argue that AI advances are (Gmyrek et al., 2023). relevant to the evolving job market,
unlikely to increase inequality as fostering inclusive economic
much as previous automation The impact of AI on jobs in growth through investments in
technologies because their impact advanced economies could education and infrastructure, and
is more equally distributed across exacerbate existing inequalities reevaluating taxation policies
demographic groups, but there and make it necessary to develop to ensure a fair distribution of
is also no evidence that AI will substantial adaptation strategies. the gains generated by AI. In
reduce labour income inequality AI has the potential to reshape addition, promoting innovation
(Acemoglu, 2024). income distribution by decreasing and entrepreneurship among
the labour share and increasing marginalized communities could
Advanced economies, particularly the returns on capital. As AI help to mitigate the adverse
those with high levels of automation and automation become more effects of AI-induced income
and technology adoption, are integrated, the value created by inequality while fostering a
more likely to experience significant human labour may be diminished more equitable society.

access, transmit, store and process data and to perform AI can increase the demand for specialized
intensive computations are essential to AI development and development tools and software libraries. As the
deployment. These services include cloud computing, demand for AI models has experienced a notable surge in
which provides the necessary online infrastructure and recent years, the frameworks designed to streamline the
platforms for developing and running AI applications; AI development, testing and deployment of AI models and
model development services, which offer tools and platforms applications are also increasing. These include integrated
for creating, training and deploying AI systems; data development environments (IDEs), machine learning
services to gather, clean and label data needed to train AI libraries and AI platforms that simplify the implementation
models; and security services to protect AI systems and of AI algorithms and workflows. Software for designing
data from cyber threats. specialized AI semiconductors is also in high demand.

26
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

Figure 2.5: The AI value chain

Raw Lithography Talent and IP Data

resources

Semiconductor Assembly,
fabrication testing and
packaging of
semiconductors
AI algorithm

Memory Semiconductor
design

Computational Applications
power

Source: World Economic Forum (2024b).

Many of these goods and services are often supplied earth metals. As this industry evolves, understanding the
by a small number of economies. International trade dynamics of rare earth production becomes ever more
therefore serves as an important channel to foster AI critical. These metals and minerals are geographically
development worldwide. The production of AI technologies concentrated, with China (35.7 per cent), Brazil (17 per cent)
is heavily concentrated within a globally integrated supply and Russia (15.7 per cent) hosting the largest reserves yet
chain. As indicated in Figure 2.6, alongside the concentration discovered of rare earths (BP, 2022). As these resources
of AI models, various stages of AI production, including are essential for the production of various technologies, the
AI chip design and manufacturing, are dominated by a demand for trade in these products is expected to continue
small number of suppliers, with some critical steps to rise. Furthermore, as AI is increasingly integrated into
having fewer than three suppliers (Sastry et al., 2024). various sectors, the demand for energy to power these
This concentration augments risks to the supply chain, systems is projected to escalate further.
including vulnerabilities stemming from export controls
and potential cyber threats (Miller, 2022; World Economic AI has substantially heightened the demand for
Forum, 2023). data, fundamentally reshaping the landscape of
data usage and trade. Data assumes a dual role in the
Upstream of the value chain, trade in the extraction production of AI technology, serving both as an input and
and processing of critical metals and minerals, as an output. Given that AI systems rely heavily on data, the
well as in energy, is also likely to rise. Advanced AI demand for high-quality, diverse datasets has surged and
chips require elements like neodymium, cerium and will continue to surge. International flows of data are crucial
praseodymium. The growth of the AI chip industry is for accurate, complete and representative datasets to feed
inseparably linked to the production and supply of rare into AI systems (Aaronson, 2023).

27
CHAPTER 2: WHY IS AI A TRADE ISSUE?

Figure 2.6: Concentration of the AI chip supply chain

(percentage)

120

SAMSUNG & Intel

100
Others (mainly Intel 10%
and AMD)
5-20% Others
80 22%

60
Google Cloud
Taiwan 11%
ASML
Semiconductor
NVIDIA 100%
Manufacturing Microsoft Azure
40 80-95% Company (TSMC) 22%
90%

20 Amazon
Web Services
32%
0
Design of Semiconductor Fabrication Compute
AI chips manufacturing equipment provision

Data: 2023. Data: 2023. ASML is the only Data: 2022. This covers the entire
This excludes AI chips company capable of producing Logic chips <= 7nm cloud market and is not
that are not available Extreme Ultraviolet Lithography specific to AI compute.
to purchases. (EUV) machines.

Source: Sastry et al. (2024).

(iv) AI can reshape increase in total factor productivity, ranging between 0.55
economies’ comparative per cent and 0.71 per cent over a 10-year period.

advantages The widespread use of AI has the potential to boost

productivity significantly in both developed and
developing economies. Although the development of AI
AI can reshape comparative is likely to remain concentrated in a few large economies,
advantages by affecting productivity the cost of AI use and application in specific domains
is relatively low. This will allow developing economies to
AI is expected to enhance productivity across all leverage AI to improve productivity, enhance efficiency,
sectors in the global economy. Although the productivity access better public services and reduce costs. Examples
impact of AI is more pronounced in the services sector, include AI-driven learning systems that enable individualized
other economic sectors can also expect productivity growth. learning at relatively low cost (Muralidharan et al., 2019),
In agriculture, AI applications can be employed to forecast the use of LLMs and speech recognition software to assist
weather patterns and optimize resource management. illiterate farmers in applying for government loans (Yee,
In manufacturing, AI advancements can significantly 2023), and the implementation of AI to improve healthcare
improve efficiency by automating and optimizing routine delivery and diagnosis in Africa (Owoyemi et al., 2020).
processes and tasks, optimizing material and energy usage,
and enhancing the accuracy of predictions and forecasting
(World Economic Forum, 2022). AI can redefine the comparative
advantage of economies through
The impact of generative AI on productivity could be shifts in production dynamics
significant. McKinsey (2023) estimates that generative
AI could add the equivalent of between US$ 2.6 trillion and AI may not only enhances productivity, but also
US$ 4.4 trillion annually. Goldman Sachs (2023) estimates reshape the composition of inputs required for
that widespread adoption of generative AI could raise overall production, leading to a greater emphasis on capital
labour productivity growth by around 1.5 per cent per investment relative to labour inputs. As AI technologies
year over a decade, a similar boost to what occurred with become more advanced and widespread, businesses are
previous transformative technologies such as the electric likely to invest heavily in AI-driven automation and intelligent
motor and personal computer. However, a more recent study systems that can enhance productivity, efficiency and
by Acemoglu (2024) predicts a somewhat more moderate decision-making processes. This shift may lead to a reduced

28
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

reliance on human labour, particularly for routine and repetitive Large corporations often have extensive numbers
tasks, thereby increasing the capital intensity of production. of users and consequently vast pools of data with
which to train AI algorithms. As more users interact
This shift in production dynamics has the potential with AI systems, they generate more data, and this in turn
to reshape trade patterns. The wide adoption of AI could improves the performance and effectiveness of the AI
devalue the comparative advantage of economies abundant algorithms. This positive feedback loop enhances the value
in unskilled labour, which may lack the capability to utilize AI of the AI system for existing users, while also attracting new
effectively. In contrast, advanced economies benefiting from users, who then contribute to the growing pool of data,
higher AI intensity, driven by higher wages and capital, may setting off a feedback loop wherein dominant players
experience greater gains. Internationally mobile capital may be attract more users, generate more data and further refine
drawn towards advanced economies, leading to transitional their AI systems, solidifying their market dominance. This
GDP declines in developing economies (Alonso et al., 2022). dynamic represents a significant hurdle for newcomers
and smaller enterprises, which typically lack the resources
Conversely, new sources of comparative advantage to gather, manage and safeguard such extensive data.
may emerge from educated labour, digital connectivity Consequently, smaller competitors face increasing difficulties
and regulation. The ability to leverage AI for development in developing AI capabilities of comparable scale and
critically depends on economies’ readiness to use the sophistication (OECD, 2021; West, 2023). This may lead to
technology, which includes factors such as digital a market landscape dominated by a select few major players
infrastructure, human capital, innovation and regulation (Lee, 2024).
(Cazzaniga et al., 2024). Digital infrastructure and human
capital can be considered foundational elements of AI Several studies also demonstrate that big data and
preparedness, because they are prerequisites for its AI have resulted in industrial concentration. For
adoption. Innovation and regulation can be considered instance, Begenau, Farboodi and Veldkamp (2018) suggest
additional elements likely to influence the ability to develop AI that access to big data in finance has reduced the cost of
and maximize its economic impact. capital for large firms relative to smaller ones, leading to
increased firm-size inequality. Firooz et al. (2022) provide
As AI is energy-intensive, and many firms are seeking evidence that the development of automation technology
to decarbonize, economies with abundant renewable has contributed to the dominance of superstar firms over
energy may also have a comparative advantage. the past two decades. These findings highlight the
As noted in Box 2.1, the International Energy Agency (IEA) influence of big data and AI on market dynamics and how
estimates that electricity consumption associated with data they may consolidate power among dominant players in
centres, cryptocurrencies and AI represented almost 2 per various industries.
cent of global energy demand in 2022, and that energy
demand for these uses could double by 2026 (IEA, 2024).
To move towards net zero greenhouse gas emissions,
companies are developing strategies to rely on renewable
(v) Projection of the impact
energy for AI. Therefore, economies capable of generating of AI on trade
renewable energy may have a comparative advantage for
hosting data centres and AI infrastructure.
The WTO Global Trade Model was employed to
project the potential impact of AI on international
The development and control of trade patterns. This is a recursive dynamic computable
general equilibrium (CGE) model which enables long-term
AI technology are likely to remain projections until 2040. Based on insights from the
concentrated in large economies literature and from WTO empirical work, two sets of shocks
and companies relative to a baseline without AI were introduced, i.e.,
increases in labour productivity and reductions in trade
The substantial upfront investment in AI often results costs. It is anticipated that AI will impact trade costs through
in increasing returns to scale. AI development three main channels: improved logistics, diminished
fundamentally depends on ICT infrastructure performance, compliance costs and reduced language barriers. Four
specialized hardware and extensive data storage systems, scenarios that differ along two dimensions were considered:
all of which require substantial upfront investment. As the size of the productivity impact of AI (optimistic or
AI models advance, and their development costs cautious) and the scope for convergence between
escalate, the up-front costs of developing AI models economies and between workers with different skills
increase. For instance, training ChatGPT-3 reportedly (synergy or divergence). When combined, this leads to the
required over US$ 4 million, while GPT-4’s development four scenarios outlined in Table 2.1. Technical details on the
reportedly surpassed US$ 100 million, and the construction of the scenarios are presented in Annex 2:
operation of ChatGPT alone has been estimated to incur
US$ 700,000 per day in computer costs. The exorbitant • Optimistic global synergy: High productivity growth with
costs associated with AI development can act as barriers, universal AI adoption
hindering smaller entrants from penetrating the market, • Optimistic tech divergence: High productivity growth with
and resulting in market concentration. uneven AI adoption

29
CHAPTER 2: WHY IS AI A TRADE ISSUE?

Table 2.1: Summary of scenarios

Global synergy Tech divergence

High average global productivity High average global productivity

increase, based on Goldman increase, based on Goldman
Sachs (2023) estimate; Sachs (2023) estimate;
Uniform productivity increase Productivity increase differs by region
Optimistic across economies; according to AI preparedness;
scenario
Middle-skilled workers raise productivity High-skilled workers raise productivity
more than high-skilled workers; more than middle-skilled workers;
All regions implement trade cost Trade cost reductions account for
reductions through AI equally. regional differences in AI preparedness.
Low average global productivity Low average global productivity
increase, based on Acemoglu (2024); increase, based on Acemoglu (2024);
Uniform productivity increase Productivity increase differs by
Cautious across economies; region according to AI preparedness;
scenario Middle-skilled workers raise productivity High-skilled workers raise productivity
more than high-skilled workers; more than middle-skilled workers;
All regions implement trade cost Trade cost reductions account for
reductions through AI equally. regional differences in AI preparedness.

Source: WTO.

• Cautious global synergy: Low productivity growth with cautious scenarios, and trade cost reductions play a larger
universal AI adoption role in driving trade growth. AI is also expected to boost
• Cautious tech divergence: Low productivity growth with real global GDP by 11 percentage points until 2040 (see
uneven AI adoption Annex 2 for further details).

While high-income economies are expected to see The global trade impact of AI varies significantly
the largest productivity gains, trade cost reductions across economies and sectors. Figure 2.9 illustrates
can favour low-income economies. As shown in the projected trade changes due to AI across four income
the upper panel of Figure 2.7, productivity increases are groups: low-income, lower middle-income, upper middle-
particularly significant in higher-income economies, due income and high-income economies. The results show that,
to their greater AI preparedness and specialization in under the global synergy scenario, low-income economies
AI-intensive sectors. The bottom panel highlights that trade experience much higher trade growth compared to the tech
cost reductions can be negatively correlated with current divergence scenario, while trade growth in high-income
income level. This is particularly pronounced in the global economies remains relatively stable across scenarios.
synergy scenario, showing that lower-income economies
have more potential to reduce trade costs. Digitally delivered services are expected to experience
the highest trade growth, while other sectors will
Trade growth is projected to be highest in the optimistic also benefit. Figure 2.10 compares the projected trade
global synergy scenario, with real trade growth growth across four aggregate sectors: primary (agriculture
increasing by nearly 14 percentage points by 2040. and mining), secondary (manufacturing), tertiary digital
Figure 2.8 illustrates the impact of AI on global trade, (digitally delivered services) and tertiary other (other
comparing cumulative trade growth rates with and without services). AI is projected to benefit the digitally delivered
AI over this period. The highest global trade growth is services the most, while agricultural goods are expected to
projected for the optimistic global synergy scenario. see the smallest increase in exports. Digitally delivered
Furthermore, productivity growth and trade cost reductions services are projected to see the largest increases, with
contribute equally to trade growth in the optimistic scenarios, a cumulative growth of nearly 18 percentage points in the
whereas projected productivity increases are smaller in the optimistic global synergy scenario.

30
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

Figure 2.7: Cumulative trade cost reductions (by importer) and productivity improvements (%)
in global synergy and tech divergence scenarios

Cumulative productivity shock Cumulative productivity shock

(Global Synergy Optimistic) (Tech Divergence Optimistic)

17 17
aus eft
can
zaf kor gbr usa usa
bra e27
oas
gbr eft
15 lac 15
can
jpn aus
rus
sea
sso tur kor
min
13 row 13
chn jpn
ind e27
% change

% change
asl mex
ssp
11 11
idn
sea

oasrus
9 9
chn
zaf
min tur
bra
ind lac
7 7 mex
row
sso
idn

5 5
0 10,000 20,000 30,000 40,000 50,000 60,000 70,000 80,000 90,000 0 10,000 20,000 30,000 40,000 50,000 60,000 70,000 80,000 90,000

GDP-per capita (USD) GDP-per capita (USD)

Cumulative importer trade cost change Cumulative importer trade cost change
(Global Synergy Optimistic) (Tech Divergence Optimistic)

0 0

-2

-2
-4
% change in ad-valorem equivalent

% change in ad-valorem equivalent

-6
-4
jpn usa
-8 can
oas aus
e27 gbr
zaf
ind sea kor eft
-10 -6
chn tur oas
idn ind jpn usa
-12 mex e27 can aus
min tur
sea gbr
row bra -8 idn chn kor
rus row
-14 lac ssl min
sso sso bra eft
asl asl lac mex
-16 rus
-10
ssl
-18

-20 -12
0 10,000 20,000 30,000 40,000 50,000 60,000 70,000 80,000 90,000 0 10,000 20,000 30,000 40,000 50,000 60,000 70,000 80,000 90,000

GDP-per capita (USD) GDP-per capita (USD)

Source: Simulation results based on the WTO Global Trade Model.

Note: The figure represents the relationship between GDP per capita in 2017 and projected cumulative trade cost reduction as
ad valorem equivalents over 2018-40 and productivity increases over 2027-40 according to the optimistic global synergy and
tech divergence scenarios. Each marker represents a region. See Table A.1 in Annex 2 for a list of abbreviations of regions.

31
CHAPTER 2: WHY IS AI A TRADE ISSUE?

The expected impact of AI on real Figure 2.8: Cumulative global real trade
trade growth differs within sectors. growth rate (2023-40)
As shown in Figure 2.11, in the optimistic
global synergy cenario, digitally delivered (Difference to baseline, percentage points)
services such as education, health,
recreational and financial services, as 14 13.6
well as manufacturing sectors, such 12 9.3
10 10.2
as processed food, are projected to 6.9
8
experience significant trade growth, 6
largely driven by trade cost reductions. 4
Conversely, sectors related to natural 2
resource extraction (e.g., petroleum and 0
Tech Tech Global Global
oil) and manufacturing sectors, such as Divergence Divergence Synergy Synergy
textiles and computer, electronic and Cautious Optimistic Cautious Optimistic
optical products, are expected to see
Trade costs Productivity growth
limited growth due to AI.
Source: Simulation results based on the WTO Global
Trade Model.
Note: This figure demonstrates the impact of policy shocks
on projected cumulative global real trade growth (in percentage
points) over the period 2023-40 across four policy scenarios.
The values represent deviations from the baseline scenario.
The values above the bars indicate the total effect.

Figure 2.9: Cumulative regional real export growth rate (2023-40)

(Difference to baseline, percentage points)

Low-income 18.1 Lower middle-income

18 18 16.1
16 16
14 12.0 14
12 10.4 12 10.9 10.4
10 10
8 6.5 8 6.9
6 6
4 4
2 2
0 0
Tech Tech Global Global Tech Tech Global Global
Divergence Divergence Synergy Synergy Divergence Divergence Synergy Synergy
Cautious Optimistic Cautious Optimistic Cautious Optimistic Cautious Optimistic

Upper middle-income High-income

16 14.5 16
14 14 12.6
12 10.4 10.5 12 10.2
10 10 8.3
7.5 6.8
8 8
6 6
4 4
2 2
0 0
Tech Tech Global Global Tech Tech Global Global
Divergence Divergence Synergy Synergy Divergence Divergence Synergy Synergy
Cautious Optimistic Cautious Optimistic Cautious Optimistic Cautious Optimistic

Trade costs Productivity growth

Source: Simulation results based on the WTO Global Trade Model.

Note: This figure demonstrates the impact of AI on projected cumulative regional real exports growth (in percentage points)
over 2023-40 in four policy scenarios. The values above the bars indicate the total effect, compared with the baseline scenario.

32
 CHAPTER 2: WHY IS AI A TRADE ISSUE?

Figure 2.10: Cumulative sectoral real exports growth rate (2023-40)

(Difference to baseline, in percentage points)

Primary Secondary
18 18

16 16

14 14 13.2

12 11.0 12
9.9
10 10 9.1
8.4
7.5
8 8 6.8
5.9
6 6

4 4

2 2

0 0
Tech Tech Global Global Tech Tech Global Global
Divergence Divergence Synergy Synergy Divergence Divergence Synergy Synergy
Cautious Optimistic Cautious Optimistic Cautious Optimistic Cautious Optimistic

Digitally delivered services Other services

18
17.8 18

16 16 15.7
14.0
14 14
12.1
11.4 10.9
12 12
9.5
10 10
7.9
8 8

6 6

4 4

2 2

0 0
Tech Tech Global Global Tech Tech Global Global
Divergence Divergence Synergy Synergy Divergence Divergence Synergy Synergy
Cautious Optimistic Cautious Optimistic Cautious Optimistic Cautious Optimistic

Trade costs Productivity growth

Source: Simulation results based on the WTO Global Trade Model.

Note: This figure demonstrates the impact of AI on projected cumulative sectoral real exports growth (in percentage points) over
2023-40 in four policy scenarios. The yellow and blue bars represent the effects of trade cost reduction and productivity growth
respectively, and the values above the bars indicate the total effect, compared with the baseline scenario.

The varying trade projections across scenarios trade patterns through the export of intermediate goods
underscore the critical role of policy in leveraging AI and services like semiconductors and telecommunications.
for trade. A key difference between the global synergy However, this impact is not captured in the projections.
and tech divergence scenarios is the ability of developing Second, the scenario in which productivity increases more
economies to adopt AI. Developing economies that improve for middle-skilled workers than for high-skilled workers is
their AI preparedness, by enhancing digital infrastructure, inspired by arguments from scholars such as David Autor
upgrading skills, and boosting innovation and regulatory and Richard Baldwin, who suggest that AI could help
capacities, can significantly enhance their ability to leverage rebuild the middle class (Autor, 2024; Baldwin, 2024). This
AI effectively. In addition, directing AI innovation toward should ideally be grounded in quantitative analysis. Third,
benefiting the productivity of middleskilled workers could AI may lead to the substitution of labour with capital and
further help lower-income economies to close the trade and intangible assets. Although this effect is not considered in
income gap. the model, it is expected to primarily impact wages rather than
trade projections. Finally, the model assumes no emergence
While the projections are informative, several caveats or disappearance of products or tasks due to AI. However,
must be noted. First, the adoption of AI necessitates AI could lead to structural changes in the economy, creating
investments in digital infrastructure, which could affect new goods and services or rendering some obsolete.

33
CHAPTER 2: WHY IS AI A TRADE ISSUE?

Figure 2.11: Cumulative real trade growth (2023-40) in the optimistic global synergy scenario

(Compared to baseline, percentage points)

Education and human health

Digitally delivered

Recreational and other services

services

Insurance
Financial services
Communication
Business services
Other natural resources
Primary

Agriculture
Petroleum, coal products
Oil
Processed food
Other goods
Motor vehicles
Chemicals and petrochemicals
Secondary

Pharmaceuticals, rubber and plastics

Metals
Transport equipment nec
Electronic equipment
Other machinery
Textiles, wearing apparel and leather
Computer, electronic and optical products
Utilities and construction
Trade
Other services

Other services
Warehousing and support activities
Transport
Accommodation, food and services
-10 0 10 20 30 40 50 60 70 80

Productivity growth (pp.) Trade costs (pp.)

Source: Simulation results based on the WTO Global Trade Model.

Note: This figure demonstrates the impact of AI on projected cumulative sectoral real exports growth (in percentage points) over 2023-40
in the optimistic global synergy scenario, compared with the baseline scenario. The yellow and blue bars represent the effects of trade
cost reduction and productivity growth respectively; “nec” is “not elsewhere classified”. “pp” is “percentage points”.

Endnotes
1 See [Link] Section 3.1.4. Definitions used materials and products, such as algae-based advanced biofuels, synthetic
in this report are without prejudice to the views of WTO members. fabrics such as “micro-silk”, and bio-based durable packaging materials
2 The issue of the military use of AI is beyond the scope of this report. However, (Webb and Hessel, 2022).
recently, there have been various international and domestic debates, initiatives 5 See [Link]
and proposals on this matter. See for instance the Proposal for a UN General 6 See [Link]
Assembly Resolution on “Lethal Autonomous Weapons Systems (LAWS)”
(A/C.1/78/L.56, 12 October 2023) and the United States’ “Political Declaration 7 See [Link]
on Responsible Military Use of Artificial Intelligence and Autonomy” (9 November vessel-tracking/.
2023). On Lethal Autonomous Weapon Systems (LAWS), see more broadly 8 See [Link]
[Link] [Link].
weapons/background-on-laws-in-the-ccw/. 9 See the Moderator’s Report from the November 2023 Thematic Session on
3 However, as nanomaterials can also pose health and environmental challenges the “Use of Digital Technologies and Tools in Good Regulatory Practices” at the
(e.g., concerning the end of life of products containing them), the special role WTO Committee on Technical Barriers to Trade ([Link]
of regulations and policies to ensure that such risks are addressed must be tratop_e/tbt_e/tbt_0711202310_e/tbt_0711202310_e.htm).
stressed. 10 Information summarized from presentations at the WTO Committee on
4 Examples include decarbonizing carbon-intensive sectors, such as agriculture, Sanitary and Phytosanitary (SPS) Measures thematic session on digital
by optimizing production methods that reduce the emission of methane and tools on 25 June 2024 ([Link]
nitrogen oxides, as well as enabling the production of new kinds of sustainable sps_2506202410_e/sps_2506202410_e.htm).

34
 CHAPTER 3: THE POLICIES OF AI AND TRADE

3 The policies
of AI and trade
35
CHAPTER 3: THE POLICIES OF AI AND TRADE

Wealthier economies, including advanced and some

(a) AI and trade: emerging market economies, are generally better
prepared than low-income economies to adopt AI.
key policy As illustrated in Figure 3.1, both the Digital Infrastructure

considerations Index and the Human Capital and Labor Market Policies
Index — components of the IMF’s AI Preparedness Index —
are positively correlated with income levels. Higher-income
economies tend to have stronger digital infrastructure and
The discussion of how AI might reshape international more trained human capital, making them more equipped
trade raises important policy questions. The future of to adopt AI technologies.
AI and international trade hinges on the policy choices of
governments and on the strategies and priorities of industries To address the AI divide, it is crucial to invest in digital
and businesses. infrastructure to ensure that low-income economies
have the necessary technological foundation to
support AI adoption. Governments and the private sector
(i) Addressing the growing could collaborate to expand high-speed internet access,
improve electricity infrastructure, particularly through renewable
AI divide energy generation, enhance data storage capabilities, and
develop robust cybersecurity measures. Public policies need
to incentivize infrastructure development in underserved
To leverage the opportunities of AI, the digital areas, and international cooperation should focus on providing
divide between economies, in terms of both digital technical and financial assistance to developing economies.
infrastructure and skills, must be addressed. As In addition to infrastructure, bridging the AI divide requires a
discussed in Chapter 2, ensuring that workers and firms are substantial investment in human capital to equip individuals
prepared to adopt AI involves robust digital infrastructure with the skills they need to utilize AI technologies effectively.
and trained human capital. Digital infrastructure is a crucial Education and training programmes should include AI literacy,
determinant of information and communications technology coding, data analysis and other relevant skills. Public-private
(ICT) adoption, and can lay the foundation for the diffusion partnerships can play a key role in these efforts, as companies
and localized application of AI technology (Nicoletti et al., can offer practical training and resources, while governments
2020). Nonetheless, such infrastructure is of limited use can provide the necessary regulatory support, access to
without a skilled workforce capable of leveraging digital affordable devices and connectivity, and funding (see the
platforms for innovative workplace applications. opinion piece by James Manyika).

Figure 3.1: AI preparedness is higher in advanced economies

0.9
0.9 0.9
0.9
Human Capital and Labor Market Policies Index

0.8
0.8 0.8
0.8

0.7
0.7 0.7
0.7
Digital Infrastructure Index

0.6
0.6 0.6
0.6

0.5
0.5 0.5
0.5

0.4
0.4 0.4
0.4

0.3
0.3 0.3
0.3

0.2
0.2 0.2
0.2

0.1
0.1 0.1
0.1

0.0
0.0 0.0
0.0
00 20,000 40,000
20,000 40,000 60,000
60,000 80,000
80,000 100,000
100,000120,000
120,000 00 20,000 40,000
20,000 40,000 60,000
60,000 80,000
80,000 100,000
100,000120,000
120,000

GNIper
GNI percapita
capita(current
(currentUS$)
US$) GNIper
GNI percapita
capita(current
(currentUS$)
US$)

Low-income economies Emerging market economies Advanced economies

Source: International Monetary Fund (IMF) AI Preparedness Index (Cazzaniga et al., 2024). The indices are rescaled to range
between 0 and 1.

36
 CHAPTER 3: THE POLICIES OF AI AND TRADE

Opinion piece James Manyika

Harnessing technology to advance Senior Vice President of Research,
shared prosperity Technology and Society at Google;
Co-Chair of the United Nations
Artificial intelligence is the most important technology Secretary-General’s High-Level
of the present era, offering the potential to make Advisory Body on
people’s everyday lives easier, power economic Artificial Intelligence
growth, help middle-class and lower-income workers,
drive scientific and health advances, and address
longstanding development challenges. In a world
that is on track to meet only 15 per cent of the UN’s
Sustainable Development Goals (SDGs), AI provides this means providing access to AI-capable cloud
an opportunity to reverse that trendline and contribute infrastructure, computing capacity, developer tools
to progress on 79 per cent of our shared global goals and datasets relevant to AI development, while
(Hoyer Gosselink et al., 2024). equipping workers and students with foundational AI
skills that provide pathways to the modern workforce.
However, while the economic and societal opportunity Most crucially, we must place small businesses and
offered by AI is immense, we must always remember traditional industries like manufacturing and agriculture
that the benefits of new technologies are not automatic. at the forefront of AI leadership.
At this moment of excitement, it is important to take
a step back and consider the history of trade and At the international level, driving a vision of shared
technology – and make a concerted effort to build prosperity will require expanding what we think of as
an inclusive trading system around AI that avoids the “trade” – not just removing barriers to cross-border
creation of an “AI divide.” goods and services, but advancing a global strategy
to drive alignment on AI governance and security,
The combined forces of technology and trade helped support trusted data flows, enable economic integration,
lift over a billion people out of extreme poverty – an and build solutions to cross-border challenges.
achievement unparalleled in human history.1 This An inclusive AI strategy must also drive investment
dramatically changed global development, speeding in the subsea and terrestrial cables that enable
up the flow of data and enabling smaller companies participation in the modern economy (Quigley, 2024).
and economies to participate in trade. By 2016, digital
flows – often a key aspect of other global flows such At its heart, this is a modern form of capacity building,
as manufacturing, services and financial flows and with governments, industry and civil society working
other intangibles – had begun to exert a larger impact together to invest in AI infrastructure, build a global
on GDP growth than the centuries-old trade in goods resource for AI research and develop training
(Manyika et al., 2016), with a surprising share of programmes that promote AI diffusion across
the benefits from digital trade going to the services, sectors and geographies.
manufacturing and retail sectors (The White House,
2024). These trends on digital and digitally enabled In contrast, if economies cannot align trade with the
global flows have only accelerated since 2016, with mission of shared prosperity, there is a risk that AI
some estimating that up to 40 per cent of GDP now will only be adopted by wealthier economies, and
depends on global flows (Seong et al., 2022). by the wealthiest industries within those economies.
This would be harmful not just from an equity
Now with AI, economies stand on the verge of an even perspective but also from an economic perspective
more profound economic and scientific transformation – the trillions in potential economic benefits from AI
that may fundamentally shift and reshape trading are conditional on broad-based adoption of these
patterns. But it is critical to avoid creating an “AI divide” technologies, not usage by the privileged few.
(Ossa, 2023). As of 2023, 93 per cent of people in
high-income economies use the internet, compared The choice is ours. Together, let’s build a trade
with only 27 per cent of people in low-income agenda that harnesses the transformative power
economies (ITU, 2023). The UN’s AI Advisory Body of AI for all people, regardless of geographical
has rightly concluded that these divides cannot location or economic status.
persist into the AI era (United Nations, 2024a).
Disclaimer
We must work together across companies, Opinion pieces are the sole responsibility of their
governments and civil society to harness AI to advance authors. They do not necessarily reflect the opinions
a vision of shared prosperity. At a national level, or views of WTO members or the WTO Secretariat.

37
CHAPTER 3: THE POLICIES OF AI AND TRADE

literature’s influence.3 As illustrated in Figure 3.3, the affiliation

Figure 3.2: Number of AI patent filings by of teams researching and publishing on AI systems that
geographic regions, 2013-22 “demonstrate the ability to learn, show tangible experimental
results, and contribute advancements that push the
140,000 boundaries of existing AI technology” showed an important
switch in the second part of the 2010s: whereas most of
the research was led by academia prior to 2016, industry
120,000
took the lead in the number of publications afterwards.

100,000 Investment in AI is accelerating rapidly, with the

United States leading in private investment. AI funding
stems from a variety of sources, including private companies,
80,000
venture capital firms, government funding, academic
institutions, corporate partnerships and angel investors. The
60,000 United States leads in terms of total AI private investment.
In 2023, the US$ 67.2 billion invested in AI in the United
States was roughly 8.7 times greater than the amount invested
40,000
in the next highest country, China (US$ 7.8 billion), and
17.8 times the amount invested in the United Kingdom
20,000 (US$ 3.8 billion). Since 2022, the United States has
experienced a notable increase in private investment in AI
(22.1 per cent) (Maslej et al., 2024).
0
2013

2014

2015

2016

2017

2018

2019

2020

2021

2022

The disparity in private investment in AI becomes

particularly pronounced in generative AI. Despite a
Australia European Union and Republic of Korea
United Kingdom recent decline in overall AI private investment, funding for
China India United States generative AI has surged, reaching US$ 25.2 billion in 2023
Canada Japan Other (Maslej et al., 2024). However, this surge is heavily
concentrated in a few economies, with the United States
Source: Centre for Security and Emerging Technology (2024).
taking the lead. In 2023, the United States surpassed the
combined investments of the European Union plus the United
Kingdom in generative AI by approximately US$ 21.1 billion.
Venture capital investments in generative AI have also been
In addition to differences in the ability to adopt AI led by the United States, with a steep jump in 2023 to over
technologies, the AI divide across economies, reflected US$ 16 billion going towards generative adversarial networks
in AI research and development (R&D), investment (machine learning models that generate new data mimicking a
and expertise, highlights the need to address gaps in given dataset) for AI training and generative AI for text, image
AI capabilities. There is a significant divide between and audio.4
economies leading in AI R&D and the rest, especially
developing economies and least-developed countries The demographics of professionals with AI skills are
(LDCs). As illustrated in Figure 3.2, China is by far the leading largely male and located in Europe and North America.
economy in terms of the number of patents registered, with According to a developer survey by Stack Overflow, a
86,663 AI patent applications and 15,869 patents granted question-and-answer platform for programmers,5 94.24 per
in 2022, followed by the Republic of Korea and the United cent of data scientists and machine learning professionals are
States.2 This disparity reflects the underlying technological male, and the majority are located in Europe and North America
differences between economies and underscores the ([Link], 2024). Moreover, data from the Computing
importance of facilitating technology dissemination and Research Association,6 although limited to the United States
technical assistance to bridge the gap globally. There is a and Canada, reveal that the representation of women among
marked division between where the research, patents and new AI and computer science PhDs has remained stagnant at
investments in AI are located and where they are lacking, and approximately 20 per cent since 2010. This persistent gender
there is a growing risk of further exacerbating this division, gap underscores an ongoing challenge within the field.
which exists both between and within economies and
between urban and rural, less digitally connected areas. This imbalance may be further exacerbated by the
race to nurture AI through government subsidies.
The number of published articles on AI has increased As discussed in Chapter 3(b), a number of governments
steadily, with industry taking the lead. The number of are launching domestic initiatives to promote AI, backed by
published articles on AI has increased steadily, except in generous state support. However, as most of this support is
the United States which saw a drop in 2022. Although being provided by high-income economies, it may exacerbate
China is leading in terms of the volume of published scholarly the AI divide among economies. The relative concentration
articles, it is worth noting that the United States ranks of AI supply chains can also result in trade imbalances in
first in terms of the number of citations, an indicator of the AI-related goods and services.

38
 CHAPTER 3: THE POLICIES OF AI AND TRADE

Figure 3.3: Affiliation of publication research teams, 1950-2022

120

100

80

60

40

20

0
1950
1952
1954
1955
1956
1957
1959
1960
1961
1962
1968
1970
1974
1975
1976
1977
1979
1980
1981
1982
1983
1984
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
Academia Academia and industry collaboration Industry Other

Source: Epoch (2024).

Beyond the digital divide across economies, industrial advantage. Traditional antitrust frameworks may struggle to
concentration is prevalent in AI within economies adapt to the dynamic nature of AI-driven markets, requiring
due to increasing returns and network effects. competition authorities to develop new analytical tools,
As discussed in Section 2(b)(iv), as the development of AI data access mechanisms and regulatory frameworks to
models progresses and their development costs escalate, effectively safeguard competition and consumer welfare in
only large firms can afford the substantial up-front investments the AI era (see the opinion piece by Shin-yi Peng).
required. This creates a significant barrier for newcomers and
smaller enterprises, making it difficult for them to compete. There is growing scrutiny of mergers in the AI market
The high initial costs of developing cutting-edge AI models and and growing interest in better understanding the
the necessity for extensive data and computational resources implications of AI on competition.8 Traditional antitrust
further consolidate the dominance of established players. policies, which apply after the fact, when market competition
has already been impacted, are slow and focus on prices,
The widespread adoption of AI in markets can and are not sufficient to address competition issues
heighten the risk of collusion between companies. raised by AI. The competition challenges raised by AI have
AI systems integrated into pricing strategies and market led to renewed calls for a collective international approach
analysis can enable companies to monitor competitors’ to regulation and for enforcement of competition in
pricing behaviour and adjust their own prices accordingly. digital markets.
While this may optimize profits individually, it can collectively
lead to tacit agreements or collusion among competitors to
maintain higher prices (Assad et al., 2024; OECD, 2021a).
Moreover, AI’s ability to process vast data and predict market
(ii) Preventing further
trends may enhance firms’ coordination in pricing strategies, digital trade barriers
exacerbating market concentration.7

The special features of AI present challenges for Cross-border data flows are essential to AI.
competition authorities. The opacity of AI algorithms and As discussed in Section 2(a)(i), amassing vast datasets
the sheer volume of data they process can obscure anti- is vital in order to train algorithms, and data flows are
competitive practices such as price collusion, exclusionary integral to the real-time use of AI technologies. Breadth and
behaviour and discriminatory practices. Moreover, AI-driven variety of data are as important as volume.9 For AI to be
mergers and acquisitions may raise concerns about effective and deliver accurate predictions that are not
market dominance and barriers to entry, as algorithms susceptible to bias and discrimination, algorithms need to
and data assets become pivotal assets for competitive be built on high-quality, accurate and representative data.

39
CHAPTER 3: THE POLICIES OF AI AND TRADE

Cross-border data flow restrictions can negatively firms, but may do so at the expense of overall quality,
impact AI innovation and development, and can thereby undermining innovation and the full potential of AI
increase costs for firms. Such restrictions have a general (Goldfarb and Trefler, 2018).10 Cross-border data flow
negative impact on productivity, economic growth and restrictions also impose extra costs on firms wanting to do
innovation domestically and globally (Aaronson, 2019; business internationally. A recent study on the implications
Goldfarb and Tucker, 2012; Luintel and Khan, 2009; of data flow restrictions on global GDP and trade finds
Maskus and Reichman, 2004; OECD, 2016), but are of that if all economies fully restricted their data flows, it could
particular concern for AI innovation and development. result in a 5 per cent reduction in global GDP and a 10
Because AI requires vast amounts of good quality data per cent decrease in exports (OECD and WTO, 2024).
in order to be trained, and this often involves merging different To comply with data flow restrictions, firms may need to
data sources together, cross-border data flow restrictions establish a presence and duplicate activities across various
are likely to affect the quality and accuracy of AI models jurisdictions and devise a system to ensure that data are
and the scalability of AI applications significantly. By not routed internationally. While technically feasible, doing
limiting the ability of foreign firms to access data from a this can be particularly costly, especially for small businesses
given jurisdiction, such measures could favour domestic (Goldfarb and Trefler, 2018).

Box 3.1:
AI and consumer protection
AI opens significant opportunities In addition, while AI technologies can to harness algorithms and protect
for consumers, but also increases contribute to effective moderation consumers is being discussed in
the possibilities of covert influence, for the benefit of consumers, they various jurisdictions, including the
raising significant concerns over can also deliver inaccurate, biased European Union12 and the United
the exploitation of personal or discriminatory responses that Kingdom (Holmes, 2024).
information and violation of privacy, can also harm consumers.
manipulation and disinformation. However, national approaches
Finally, sellers may not fully take do not adequately protect
For consumers, AI can provide into account potential harm consumers in the case of cross-
major benefits, such as caused to consumers as a result border transactions (Jones, 2023).
individualized recommendations of consumer data misuse due to Obtaining redress in case of harm
and time-saving (e.g., AI voice the difficulty in tracing that harm remains particularly challenging
assistants can order groceries back to the original data collector. in the event of international
instantly, saving consumers Consumers may not, therefore, transactions. Although some level
hours of shopping time). The challenge data use after the data of international collaboration and
ability of AI models to establish is collected (Agrawal et al., 2019). regulatory discussions exists among
correlations between consumers’ national bodies, this cooperation
data and possible responses to Notwithstanding the fact that remains fragmented and does not
advertisements in order to predict traditional consumer protections establish an effective, transparent
consumers’ behaviour provides firms laws may apply to most scenarios framework for enforcing consumer
using AI with the unprecedented of AI use cases, often providing rights across borders (Goyens,
ability to trigger specific reactions adequate legal remedies without 2020), leading some experts to
through individualized advertisements the need for new regulations, call for new forms of international
and communications. However, this measures to regulate algorithmic regulation and cooperation to
can also exacerbate asymmetry of harm and protect consumers have protect consumers, especially
information between companies emerged in recent years in various against AI harm (Jones, 2023).
and consumers, and can lead to jurisdictions. Under the EU General
manipulation and exploitation of Data Protection Regulation, for Besides potential violations of
consumer behaviour. example, individuals have the right privacy and personal integrity,
to contest decisions made by disinformation and manipulation,
The use of algorithms to fix prices algorithms, request human oversight, the difficulty of assessing the
can lead to price efficiencies and withdraw from personalized safety and security of AI-enabled
passed on to customers, but can advertising driven by algorithmic products and services adds to
also be used to exploit consumers’ methods. China has also developed the complexity of protecting
willingness to pay a certain price comprehensive regulations to govern consumers in an AI-driven age
in the interest of the firm. algorithm use. Further legislation (see Chapter 3(a)(iii)).

40
 CHAPTER 3: THE POLICIES OF AI AND TRADE

Opinion piece Shin-yi Peng

AI: Amplifying the digital trade issues Distinguished Professor of Law,
National Tsing Hua University
Many of the challenges brought by digital technologies
that the WTO has faced in the past decades are
now amplified by AI, including issues associated with
classification, non-discrimination, data governance and Data governance in the age of AI requires perspectives
competition policy. that safeguard social values, including privacy, security,
free speech, cultural expression and algorithmic ethics.
First of all, AI acts as a facilitator for complex products
that bundle goods and services, which calls for further Finally, competition authorities worldwide are increasingly
thinking about how to adjust the goods/services legal taking or considering approaches that impose additional
silos under the WTO to address issues stemming from obligations on AI-powered big tech companies. Their
the merging of physical and digital realms. It seems likely potential abuses of market power, including self-
that the goods/services dichotomy in applying trade rules preferencing practices that promote their own services
will increasingly trigger new levels of inconsistency and within search results, algorithmic cartels and other
legal uncertainty. cross-border collusive arrangements, can be more
meaningfully addressed through competition disciplines
Second, more and more AI-based services will be able to at the international level. To what extent does algorithmic
compete directly with or substitute human professionals. practice constitute a trade barrier to goods or services,
Questions such as to what extent automated legal advice and how do anti-competitive market concentrations
tools and human attorneys should be considered to exclude foreign suppliers from a market? The reactivation
be “like services suppliers” may emerge sooner than of the WTO Working Group on Trade and Competition
expected. It remains to be seen how far concepts such as is more urgent than ever. If a set of general or sector-
“technological neutrality” or “evolutionary interpretation” specific competition disciplines could be established
can serve to clarify the scope of the GATS commitments at the WTO, it would be less necessary for competition
of market access and national treatment. authorities in developing countries and LDCs to enforce
competition law after the fact.
Third, AI presents new challenges for data governance.
Digital platforms’ advertising algorithms, or, more Disclaimer
generally, their overall business models, intensify the Opinion pieces are the sole responsibility of their
perils associated with the data-driven economy. WTO authors. They do not necessarily reflect the opinions
rules can play a more active role in reducing such perils. or views of WTO members or the WTO Secretariat.

However, the large datasets required by AI models Restrictions on cross-border data flows also
raise significant privacy concerns. AI introduces new negatively impact trade in AI-enabled products. While
privacy issues for individuals and consumers, leading there is empirical evidence that AI significantly enhances
to a trade-off between the necessity of accessing large international trade in digital services, cross-border data
amounts of data to train AI models and privacy protection. regulation can impede such trade. Sun and Trefler (2023)
The continuous tracking and profiling of individuals’ online find that restrictions on data flows can reduce the value of
and offline interactions by AI algorithms raise significant AI-enabled apps, making them less attractive to international
concerns about data privacy, consent and control over users. While AI leads to a 10-fold increase in the number of
personal information. Furthermore, as AI algorithms become foreign users, the impact of AI on foreign users is halved if
increasingly sophisticated in their ability to infer insights the foreign users are in an economy with strong restrictions
and predict behaviours based on user data, there is a on cross-border data flows. Thus, economies with strict data
pressing need for robust privacy regulations, transparent regulations may lose out on AI-driven trade opportunities.
data practices and enhanced user control mechanisms Striking the right balance between protecting privacy and
to safeguard individuals’ privacy rights and ensure fostering innovation is therefore crucial for maximizing the
ethical and responsible AI deployment.11 AI also benefits of AI for international trade. However, cross-border
introduces new privacy concerns for consumers data flow measures, when aimed at protecting privacy, can
(see Box 3.1), and the use of data as inputs into AI help to build trust in AI systems and promote their wider use.
models also raises IP concerns (see Section 3(a)(iv)). A study by OECD and WTO (2024) on the implications of
As a result, a delicate balance needs to be found between data flow restrictions finds that, although removing data flow
privacy concerns and the need to access large amounts of regulations across all economies would reduce trade costs,
data to train AI models (see also the opinion piece by it would also undermine trust, leading to reduced consumer
Shin-yi Peng). willingness to pay for products and a negative effect on GDP.

41
CHAPTER 3: THE POLICIES OF AI AND TRADE

(iii) Ensuring the which has been described as presenting two dimensions,
legal and technical. Addressing the legal dimension requires
trustworthiness accessing the source code. This may prove difficult as source
of AI without codes are normally proprietary, i.e., protected by IP, normally
in the form of trade secrets. There are, however, regulatory
hindering trade ways to deal with this challenge, for instance, by allowing
forced source code disclosure for regulatory or law
enforcement purposes,18 even if in practice this may not be
Standards and technical regulations play a key role easy or warranted.19 To some, the technical dimension of the
in ensuring that AI is trustworthy and, through this, black box problem may be even more significant, as the
in promoting trade in AI-enabled products. There is opacity of an AI system may persist even when access to
growing consensus concerning the pivotal role that the source code is free or has been voluntarily or mandatorily
regulations, standards and other government interventions granted. Indeed, there may be instances when AI applications
can play in ensuring that AI is trustworthy, i.e., that it meets are so complex that even programmers themselves are not
expectations in terms of criteria such as reliability, security, able to divine an intelligible explanation from the source code
privacy, safety, accountability and quality in a verifiable and other proprietary information and data as to why and
way.13 Ultimately, this means striking a regulatory balance, how certain decisions and classifications were reached by
whereby the benefits of AI are harnessed while its risks are the AI system. For some, this means that, until this technical
mitigated. Ensuring trustworthiness is not only important for challenge is satisfactorily addressed, regulatory solutions
what happens within economies. It is also relevant for what based on open-source disclosure may be “significantly
happens outside economies and between borders. Indeed, frustrated” (Lin, 2021; Mitchell et al., 2023; Pasquale, 2015).
the internal regulations that governments adopt to protect
their consumers can help to build consumers’, importers’ and Adding to the difficulty in pinpointing the source
other stakeholders’ trust in AI-enabled products, thereby of vulnerability of an AI-enabled product is the fact
fostering trade in such products. that their evolving nature may be also triggered by
external factors. Such factors include customization: the
Striking a balance between regulating AI for legitimate ability of millions of individuals to “personalize” their AI-
policy reasons and enabling trade to flow as smoothly enabled products in almost infinite different ways, posing
as possible can be particularly challenging. While the a challenge for regulators to anticipate potential risks
challenge of striking the right balance between regulation and associated with each unique customized product. Another
free trade is not new, AI’s evolving, opaque and multifaceted factor is connectivity, which may render products vulnerable
nature, and the new types of risks associated with it, are making to cyberattacks or cyberthreats by bad actors that can be
this balancing act in AI regulation and governance particularly located anywhere in the globe. These factors further increase
complex (see also the opinion piece by Eduardo Paranhos). the difficulty for regulators in anticipating and addressing
a wide range of possible unforeseeable and unintended risks
Regulating AI requires regulating a product’s over the lifecycle of these products (Lund et al., 2023).
“behaviour”. As mentioned in Chapter 2, “autonomy” is one
of the unique attributes of AI. The fact that AI systems can AI’s dual-use potential may add another layer of
imbue products with various degrees of “autonomy” means complexity. As noted in Chapter 2(a), AI’s dual-use nature
that they may generate new forms of risks stemming, not from means that it can be employed for both civil and military
problems related to the physical components of the product, purposes. This may add a domestic security and geopolitical
but instead from the way AI can make the product “behave”.14 dimension to AI’s governance, making regulatory interventions
Such risks are not easy to foresee, control or even quantify.15 and cooperation even more complex (Csernatoni, 2024;
As Judge et al (2024) note, a unique, defining technical Klein and Stewart, 2024; Pouget, 2023; Raul and Mushka,
characteristic of AI is that, unlike all other engineered systems, 2024). A related issue concerns policy and regulation in
AI’s “behaviour” is not dictated or pre-determined by its the area of AI and cybersecurity (see Box 3.2).
programme code; it is an “emergent” property. Therefore, AI-
enabled products may generate risks for reasons other than For goods, “traditional” regulations and standards that
those inherent to the tangible elements in the products normally focus on tangible, visible, static product
themselves. For instance, some consider that the “behaviour” requirements may not be able to address risks
of AI-enabled co-bots (i.e., collaborative or companion stemming from the integration of AI into “traditional”
robots), if unchecked, could provoke mental health problems products. The changeability of AI-enabled products, resulting
in the humans they accompany.16 from the evolutionary nature of AI, makes regulation a
perennial moving target. AI systems confer new properties
The opacity of the behavioural nature of AI can make and functions to the products into which they are embedded.
regulation even more challenging. Risky “behaviours” of As stressed in Chapter 2, these products’ properties and
AI-enabled products may be linked to the way their algorithms functions can be described as “dynamic”, i.e., they change
are designed. AI algorithms are notoriously opaque (Lim, 2021; overtime as a consequence of constant changes occurring
Lund et al., 2023). As noted in Chapter 2, transparency and throughout the AI system’s lifecycle via software updates
explainability are critical for understanding how and why AI or other self-improvements resulting from the algorithmic
systems work and behave the way they do.17 This challenge “learning” process. This contrasts with the “static” properties
is commonly referred to as the AI “black box” problem, of more traditional products, which normally remain

42
 CHAPTER 3: THE POLICIES OF AI AND TRADE

Box 3.2:
AI, cybersecurity and technical barriers
to trade (TBT)
AI’s ability to analyse large datasets adopting cybersecurity-related raised in the last three and a half
can help in countering cyber measures and policies, many in years alone.
threats and responding to malicious the form of TBT measures, i.e.,
cyber-attacks. However, there are technical regulations, standards Cybersecurity was the focus, for
concerns related to potentially and conformity assessment the first time, of a specific thematic
biased decision-making, the lack of procedures. session of the TBT Committee
transparency and explainability of organized in 2023. Given the
AI systems, and potential misuse Indeed, cybersecurity-related global nature of the problem, it was
or abuse. Bad actors can use AI to TBT measures have recently argued in that session that unilateral
create new malware, to design new, become one of the most prominent government interventions in this
sophisticated, or targeted phishing digital-technology-related issues area should be avoided, as they
attacks, to identify new avenues of discussed in the WTO TBT could ultimately undermine global
attack, and to create deep fakes. Committee. To date, more than cybersecurity efforts. The need for
Unsurprisingly, cybersecurity is a 90 cybersecurity-related TBT governments and the private sector
core concern expressed not only measures have been notified to to work in a more coordinated and
in domestic AI policies but also the Committee, around 65 per collaborative manner to address
in international AI principles and cent of these in the last three and rising regulatory fragmentation
governance discussions. a half years. Members have also and divergence in this area and
increasingly raised specific trade find better ways to fight increasing
Cybersecurity vulnerability risks are concerns (STCs) in the cybercrime and cyber incidents was
growing as digital technologies are TBT Committee against also underscored. In this respect,
permeating more and more societies cybersecurity-related TBT efforts to develop ambitious, fair and
and economies. In response, measures: of the 29 STCs raised inclusive cybersecurity international
governments are increasingly since 1995, 38 per cent were standards were highlighted.

essentially the same throughout their lifecycle. Many of the For instance, in an AI-enabled autonomous vehicle, mechanical
constant changes to properties and functions in AI-enabled malfunctions and/or algorithmic flaws in its internal and
products are meant to be beneficial improvements (some external cameras can both cause injuries (material) and affect
even call this “evolution”).20 However, this dynamic process the privacy of passengers or pedestrians (immaterial).21
means that known risks and concerns may also be constantly Likewise, product specifications laid down in one-size-fits-all
changing, or new ones may be emerging. For AI-enabled regulations and standards may be ill-suited for regulating AI-
products, as is the case for most other products, specifications enabled products with different customized solutions (Lund
and requirements will continue to be needed to address et al., 2023). To address such regulatory challenges, while
risks associated with their “physical” aspects (e.g., hazards supporting the deployment of, and trade in, trustworthy AI-
from defective mechanical components of an autonomous enabled products, it has been proposed that regulators think
vehicle). However, for some, such “traditional” specifications of creative ways to ensure that product requirements and
and requirements may be ill-suited or insufficient to address specifications are dynamic and adaptable to the behavioural
situations where the root cause of a risk is not a mechanical and evolutionary nature of these technologies, to ensure that
or “physical” failure, but an algorithmic design flaw or they do not become obsolete as AI characteristics, risks
other problem with the AI system embedded in the and vulnerabilities evolve throughout the product lifecycle.22
product and which may cause it to display risky “behaviour”
(e.g., an autonomous vehicle that causes injuries to people or The constantly evolving nature of AI-enabled products
damage to property). may also necessitate new approaches to certify their
compliance with regulatory requirements. Indeed, if
AI-enabled products may cause not only material but an AI-enabled product has successfully undergone testing,
also immaterial risks. An AI-enabled product may present verification or other certification procedures prior to being
both material risks, which are easy to quantify and measure placed on the market, this may not necessarily mean that
(e.g., physical injuries or damage) and immaterial risks the product will remain certifiable throughout its lifecycle.
(e.g., privacy or other fundamental rights), which are more AI-enabled products, in particular internet-connected IoTs
difficult to quantify and measure. Material and immaterial or robotics, may generate new risks after their deployment
risks can sometimes even stem from the same situation. due to “mutability” factors such as new updates, new data,

43
CHAPTER 3: THE POLICIES OF AI AND TRADE

unforeseen changes of attributes and functions due to difficult to separate value judgements from technical detail”.
user customization, or unforeseen autonomous behaviours Some have even questioned whether this could ever be done
(see Box 3.2).23 As already discussed, assessing the in practice.26
conformity of some AI-enabled products with underlying
technical regulations and standards may also require Such non-typical or immaterial AI-triggered risks and
access to source code, which raises IP-related issues concerns may also be intrinsically prone to regulatory
(see Chapter 3(a)(iv)). Regulators may also face challenges fragmentation, which could hinder trade. Indeed, it might
in assessing the compliance of AI-enabled products with be difficult for legislators to agree on common international
various novel regulatory requirements that aim, for example, denominators with respect to some AI-related societal values
to assess the quality of data used in such products. In and concerns such as ethics, privacy or human rights, the
light of such a multiplicity of challenges, some consider relative importance of which may vary across economies.27
that regulators may need to re-evaluate their conformity Unnecessary or avoidable regulatory fragmentation could,
assessment approaches and come up with methods of in turn, hamper the opportunities and benefits associated
ensuring effective continuous compliance of ever-changing with AI (Bello Villarino, 2023; OECD, 2022a). In particular,
AI products with underlying technical regulations and it could result in high regulatory compliance burdens and
standards (Lund et al., 2023; Meltzer, 2023).24 costs, and consequently create non-tariff barriers to trade for
AI businesses.
The integration of AI in goods and services has
also broadened the scope, number and nature of
risks and concerns that regulations and standards
need to address. As mentioned above, in addition to
(iv) How AI is shaped by
“traditional” regulatory concerns, such as interoperability, and may reshape IP
safety, security, quality, and the protection of human life or
health, the use and deployment of AI may also create
various “non-typical” risks, that some even qualify as AI poses new conceptual challenges for the
“existential” (UNDRR, 2023), and may raise complex ethical traditional, human-centric approach to IP rights.
and societal questions affecting public morals and human Balanced IP rights and their enforcement have an important
dignity.25 If AI is trained on biased and skewed datasets, role to play in ensuring both equitable access to AI technology
it may perpetuate or exacerbate biases or discrimination and a fair distribution of economic gains from its use.
against minority groups and infringe upon individual rights AI raises several important questions in this respect.
and freedoms (see Chapter 2). AI-enabled goods and
services are also a cause for significant concern with regard A first question concerns what form of IP protection
to data privacy, as they involve the collection, processing AI algorithms are granted. If the IP protection is based
and storage of vast amounts of user data (see Chapter 3(a)). on the fact that these algorithms are trade secrets – and
In addition, as mentioned above, AI is a technology prone thus that secrecy is an essential requirement on which to
to dual use, which may raise complex geopolitical and establish IP protection – this raises issues concerning a lack
domestic security issues and lead to further regulatory of transparency. Alternatively, new and inventive algorithms
fragmentation. Finally, both AI “inputs” and “outputs” raise may be protected by patents in some jurisdictions, with
new and complex issues of IP protection and ownership the patent system’s mandatory disclosure mechanism
(see Chapter 3(a)(iv)). yielding extensive information about AI technologies, which
directly passes into the public domain in many economies
These concerns render it challenging to design proper (WIPO, 2024). However, patent protection may constrain
regulatory solutions to ensure the trustworthiness of development of algorithms in economies in which patents
and support trade in AI and AI-enabled products. As have been taken out. Copyright, another type of IP
already mentioned, AI raises societal and ethical concerns protection, can be automatically extended to both source
(“immaterial” risks) that, unlike “traditional” concerns such as and object code, which may constrain analysis and
health and safety (“material” risks), are not typically a subject use of algorithms. As envisaged by the objectives of
for technical regulations and standards. Such “non-technical” the international IP system, appropriate exceptions and
concerns are more difficult to regulate, monitor and enforce limitations to IP rights protection are needed to balance
compared to more traditional regulatory objectives, such the different interests and to ensure appropriate access
as product safety or the protection of human health or life, and dissemination of AI technology. These regulatory tools
which can be addressed in more “technical” and objective may have to be adapted for this specific context, and some
ways. It has been argued that AI governance and regulatory jurisdictions have taken legislative steps or developed
frameworks may require norms, regulations and standards policies to encourage the development of open-source
that are perhaps better described not as purely “technical”, AI technologies.28
but instead as “socio-technical” instruments, i.e., combining
technical issues with broader societal considerations A second question concerns the use of copyright-
(Dentons et al., 2023; Kerry, 2024; Meltzer, 2023). Pouget protected data as AI inputs. Under the current
(2023) argues that developing socio-technical regulations international IP legal framework, materials such as original
is challenging in situations where both the technology and texts, images and compilations of data may be subject to
the harms it can cause are “so complex that it becomes copyright protection. This may raise the question of whether

44
 CHAPTER 3: THE POLICIES OF AI AND TRADE

Opinion piece Eduardo Paranhos

Navigating AI regulation: Lawyer, LL.M – University of
balancing innovation, risks and London (LSE); Chevening Scholar;
regulatory defragmentation Head of AI Working Group at Brazilian
Software Association (ABES)
It is not a simple task to determine when a new social
or economic phenomenon warrants regulation. Those
challenges can be further amplified when the new relation to privacy, product safety, consumer protection
scenarios take the form of innovative technologies, and internet regulations, as well as some downstream
posing both risks and opportunities. In addressing this, regulations, such as Brazil’s health authority
regardless of the nature of the changes, it is important ordinance on software as a medical device (SaMD).29
to reflect on a few foundational questions: (i) what risks For instance, a data breach that occurs as a result
and opportunities are at stake; (ii) how well understood of using an AI system is subject to the same controls
those new technologies are, so that the tools to tackle and remedies provided for in Brazil’s privacy law as
the possible risks can be properly balanced; and (iii) other breaches that occur without the use of AI.
which aspects of the technological progress indeed The privacy law also covers the potential misuse of
require new rules, vis-à-vis the existing laws. sensitive personal data that generates biased outputs
in the same way as similar misuse in offline settings.
AI is transforming the way we work, communicate
and create content faster than ever before. Fostering However, it is of pivotal importance to map the gaps
the development and implementation of AI solutions in the current legislation clearly, so that fresh AI
has the potential to increase efficiency and job regulations can address that very gap, avoiding
quality, as noted, for example, in recent studies by the overlaps and the resulting legal uncertainty.
International Labour Organization and the consultancy Another critical point which could make prescriptive
firm McKinsey. Yet, for these benefits to materialize, models problematic is the emphasis on regulating the
we should consider which traits of the AI systems “development” of AI, instead of focusing on high-risk
should really be regulated and aim to establish a “uses” of the technology.
model that, at the same time, is capable of protecting
society and promoting – rather than discouraging – One final consideration refers to the level of
research and development. preparedness that the upcoming regulations should
display to be able to evolve with the technology. It
It is an oversimplification to say that economies are seems more realistic that context-based regulations
mostly weighing up either context-principle based should gradually progress into stricter forms “if” and
formats to regulate AI, or prescriptive models with “when” needed, than the other way round. Similarly, in
a more detailed set of obligations and sanctions. the context of trade-related concerns over the adverse
effects of regulatory fragmentation, AI regulations that
In Brazil, the debates around AI regulation have so concentrate on high-risk uses – not on the development
far examined aspects from each of those possible of the technology – could facilitate the pursuit of a more
structures: a prescriptive model, openly inspired by the harmonized approach across markets. Notably, some
European regime, and another proposal for a context- economies that are global protagonists in investments
based framework anchored on widely recognized and implementation of AI30 have been leaning towards
principles for governance and risk mitigation – e.g., evolutive regulatory formats, balancing AI governance
those of the United Nations Educational, Scientific and and risk mitigation goals, while helping to raise living
Cultural Organization (UNESCO), the Organization for standards, creating quality jobs and improving
Economic Co-operation and Development (OECD) – people’s lives through responsible innovation.
reaffirming the role of existing legislation.
Disclaimer
Indeed, it is possible that most situations raising Opinion pieces are the sole responsibility of their
concerns about the deployment of AI could be dealt authors. They do not necessarily reflect the opinions
with through existing federal laws, particularly in or views of WTO members or the WTO Secretariat.

their use in training AI amounts to copyright infringement. protected materials). The application of such limited
This question translates into whether such use is now, or exceptions to generative AI, in particular, is more complex
should be, automatically permissible under exceptions to than in traditional cases, due to factors like the scale of data
exclusive copyright (e.g., for educational use of copyright- used and the purpose of the use.

45
CHAPTER 3: THE POLICIES OF AI AND TRADE

A third set of questions concerns whether AI outputs

generated autonomously can be subject to IP
protection. An AI output is created based on the patterns (b) T
 he global race
and rules learned during the training process, by means of to promote and
the input data. However, this output (including in the form
of content) is not a mere reproduction or recombination of regulate AI
this material. Rather, the output can take the form of novel and the risk of
and creative material, reflecting the AI system’s ability to
understand and mimic the complexities of human-generated fragmentation
content. Thus, AI output may encompass a wide range of
creations and innovations,31 including artwork, literary works,
music, design, films, video games and inventions. As AI is The immense potential of AI has prompted
increasingly capable of producing outputs autonomously, governments around the globe to take action to
the lines between human and AI contributions to creation promote its development and use while mitigating
or inventions are increasingly becoming blurred, making its potential risks. However, the increasing number
the question of inventorship and authorship more pressing of domestic, regional and international initiatives and
and complex. their design are fragmenting the policy landscape, with
possibly negative consequences for companies trading
Various approaches have been taken, or proposed, internationally. The economic costs of fragmentation highlight
for finding balanced and equitable answers to the importance of mitigating regulatory heterogeneity.
some of the above questions, both in terms of AI
inputs and outputs. In terms of AI inputs, proponents of
the use of copyrighted material to train AI argue that this (i) Domestic initiatives
constitutes a “transformative” use, as the model does not
replicate the copyrighted works, but instead generates
new content inspired by the learned patterns. They also Governments are using a variety of instruments to
argue that such use does not negatively affect the market promote AI and to address and mitigate its risks.
for the original works and that, in some cases, it could These range from AI-specific strategies and policies to
potentially complement that market. This is still an sector-specific legislation (including data regulations) and
ongoing debate. A balanced approach taking into account trade policy measures. However, there are already signs
both the moral and the economic interests of creators of that heterogeneity in the design of these measures may
original works and users of AI needs to be found, and the be leading to regulatory fragmentation. The sheer number
legal community continues to explore these issues. of domestic strategies and policy initiatives related to AI
Approaches to this issue differ significantly across indicates that AI is an area of priority and that a sustained
jurisdictions (see Chapter 3(b)). high level of intervention can be expected in the near future,
with a potential risk of growing regulatory fragmentation.
The question of the protection and ownership of AI
generated outputs necessitates a re-evaluation of
existing IP legal frameworks. As noted above, AI may AI strategies and policies
generate outputs in the form of a work or invention. This in
turn raises questions about whether and/or under which An increasing number of jurisdictions is putting in
circumstances IP rights can be granted for AI-generated place AI strategies and policies at the domestic
creations or innovations, and if so, who, if anyone, owns level. The number of economies that has implemented
the resulting IP, or who is liable if the output violates the AI strategies34 increased from three in 2017 (Canada, China
IP of others. Is it the developer of the AI tool, the user who and Finland) to 75 in 2023. Canada initiated the first
prompted the AI output, or neither, since the creator or domestic AI strategy in March 2017. In 2023 alone,
inventor is not human? Current IP laws attribute authorship eight new strategies were added by economies in the
and inventorship, as well as resulting economic rights, to Middle East, Africa and the Caribbean, showcasing the
humans. In some jurisdictions, AI itself is not recognized as worldwide expansion of AI policymaking (Maslej et al.,
the creator or inventor within the current IP legal framework.32 2024). In addition, or as part of domestic AI strategies,
In other jurisdictions, the issue is open to judicial interpretation governments around the world have taken over 1,000 AI
based on existing laws.33 With the widespread deployment policy initiatives.35 The majority of AI policy initiatives are
of AI, which can now be used by individuals across the concentrated in Europe, followed by Asia, the Americas and
globe, the question of protection and ownership of Africa (see Figure 3.4). Most AI-related legislation passed
AI generated outputs becomes a constant and global since 2016 aims to enhance an economy’s AI capabilities,
concern. This question not only challenges the traditional such as establishing a network of publicly accessible
understanding of creativity and ownership, but also urges supercomputers, as opposed to restrictive legislation,
a re-evaluation of existing IP legal frameworks in the age of which imposes conditions or limitations on AI deployment
AI (see also Chapter 4(f)). or usage (Maslej et al., 2024).

46
 CHAPTER 3: THE POLICIES OF AI AND TRADE

March 2024, more than a third (36 per cent) of AI policy

initiatives listed by the OECD Artificial Intelligence Policy
Figure 3.4: Share of AI policy initiatives
by region (2024) Observatory42 concerned governance of AI. Governance
aspects usually focus on the establishment of frameworks
for AI development and deployment, including vertical and
13% horizontal coordination, AI’s integration into public sector,
47% public consultation and evaluation mechanisms, and the
creation of regulatory bodies or committees to oversee
AI-related activities. Close to 19 per cent of domestic AI
23%
policy initiatives aim to provide financial support and
incentives for AI research, development and adoption.43
Around 18 per cent of these initiatives include the
development of guidelines and regulations (on issues such
as data privacy, algorithmic transparency, bias mitigation
and safety standards to promote the responsible and
ethical development and use of AI technologies), the
establishment of regulatory oversight and ethical advice
3% bodies to provide guidance and supervision in navigating
these regulations effectively, and the development of
14% standards and certification processes to facilitate the
development and adoption of AI technologies in compliance
Africa Asia Australia and New Zealand with regulatory requirements and ethical principles. Finally,
Americas Europe a large number of domestic AI policy initiatives (around
27 per cent) focuses on fostering an environment conducive
Source: OECD database of domestic AI policies to AI innovation and adoption (i.e., AI enablers). These include
([Link]
initiatives to enhance AI-related skills and education to attract
talent, public awareness campaigns to raise awareness
about AI, and the establishment of collaborative platforms,
to bring together stakeholders within the innovation
The European Union, a WTO member in its own ecosystem, and business advisory services, to support
right,36 has been particularly active, with the adoption innovation and entrepreneurship (Maslej et al., 2024). Some
of a series of policy measures to support the of these initiatives, like the 2023 US “Executive Order on the
development of trustworthy AI at the EU level.
Policy measures to support the development of
trustworthy AI include the AI Innovation Package,37 the
Coordinated Plan on AI,38 the “Proposal for standard
Figure 3.5: Share of AI policy initiatives by
contractual clauses for the procurement of Artificial
level of development (2024)
Intelligence (AI) by public organisations”, and the EU AI
Act (AIA) (European Union, 2024). The AIA, which was
formally adopted in 2024, is the world’s first comprehensive 1.1%
horizontal legal framework on AI.39 The main stated
29.42%
objective of the AIA is to ensure that AI systems within
the EU are safe and comply with existing laws on
fundamental rights, norms and values.40 The AIA adopts a
risk-based approach to regulating AI systems.41

Most domestic AI policy initiatives are implemented Category of

by developed economies, reflecting the growing country
AI divide. While a reasonable share (around 30 per cent)
of developing economies have put AI policy measures
in place, only one LDC, Uganda, has done so, with two
sector-specific policies and one general policy on AI
governance (see Figure 3.5). The increasing attention being 40.57%
paid to AI in policymaking can also be seen in references to
AI in legislative proceedings, which have increased almost
tenfold across the globe since 2016, and nearly doubled
between 2022 and 2023 (Maslej et al., 2024). Developed Developing Least-developed country

Domestic AI policy initiatives can be classified into Source: OECD database of domestic AI policies
([Link]
four broad categories: governance, financial support,
guidelines and regulations, and AI enablers. As of

47
CHAPTER 3: THE POLICIES OF AI AND TRADE

Safe, Secure, and Trustworthy Development and Use of AIA objectives is to assure the “environmental protection
Artificial Intelligence”, contain measures to support AI-related against harmful effects of [AI] systems in the Union and
hardware, such as computing infrastructure, as well as supporting innovation.”54 See also Box 2.1 on AI’s
competition and innovation in the semiconductor industry. environmental impacts.

Governments seem to be preparing or adopting an An increasing number of jurisdictions are also putting
increasing number of detailed rules and regulations in place AI-related “sandboxes”. The objective of these
related to implementing and enforcing AI legislation. is to test new economic, institutional and technological
According to Stanford University’s 2024 “AI Index”, the approaches and legal provisions under the supervision
number of AI-related regulatory measures has risen of a regulator for a limited period of time.55 About a dozen
significantly in the United States and the European Union jurisdictions, including Colombia, Estonia, the European
over the past few years. There were 25 AI-related regulatory Union, France, Germany, Lithuania, Malta, Norway, Singapore
measures adopted in the United States in 2023, including and the United Kingdom have such structures in place
three related specifically to international trade and (OECD, 2023).
international finance, compared to just one in 2016. The
total number of AI-related regulatory measures grew by Some jurisdictions are also developing “GovTech”
56.3 per cent in 2023 alone to reach 83. As for the tools (digital tools used to optimize public services)
European Union, it has passed almost 130 AI-related to address the new regulatory challenges raised by
regulatory measures since 2017, including 13 led by the AI and to promote trustworthy AI. A notable example
Directorate-General for Trade and the Directorate-General for is Singapore’s “AI Verify” tool, developed by the Infocomm
Competition (Maslej et al., 2024). Several economies are Media Development Authority and Personal Data Protection
also developing strategies or putting in place specific Commission.56 AI Verify is an open-source software tool to
initiatives to develop AI standards (see Box 3.3). assess the trustworthiness of AI systems according to a
set of criteria and factors. The tool, which is at minimum-
Environmental concerns are currently high on the viable-product stage,57 aims to automate transparency
policy agenda. Governments are therefore also starting assessment of AI systems, which would allow companies to
to draft regulatory frameworks to address the see whether new AI systems comply with relevant
potential negative environmental impacts of AI and to international standards and regulations (see the case study
harness its many benefits. For example, one of the EU’s on Singapore’s approach to AI in Box 3.4).

Box 3.3:
Domestic standards on AI
44

Standards play an important role in a variety of broad topics, most often Australia’s intention to participate
domestic AI policy approaches and those related to data management, in international standards-setting
several economies are developing quality, processing and protection, processes,50 while China’s Global
strategies or putting in place specific as well as risk management, safety AI Governance Initiative encourages
initiatives to develop AI standards.45 and security, interoperability, and international cooperation for
Some economies even recognize AI organizational governance. These developing AI standards based on
as one of the priority areas in their standards address specific technical broad consensus.51 In the same
general standardization strategies.46 requirements such as process, vein, the US Executive Order on
management and governance, Safe, Secure, and Trustworthy
As of July 2024, almost 170 measurement and test methods, Development and Use of
standards are being developed terminology, interface and architecture Artificial Intelligence mandates
or have already been published specifications, and product and relevant agencies to cooperate
by various domestic standards- performance requirements.48 with standards development
setting bodies (such as BSI, CEN, organizations to drive the development
CENELAC, NIST).47 Most of such One common feature across various of AI-related consensus standards.52
domestic standards seem to be of domestic standardization approaches
horizontal application, while others is the recognition of the importance In this respect, as in other regulatory
seem to be sectorial, i.e. only covering of engagement and cooperation on areas, domestic standardization
specific industries and sectors such AI standardization at the international efforts on AI will tend over time
as transportation, healthcare or level (Kerry, 2024).49 For instance, to rely on international standards-
energy. AI-related standards cover Australia’s AI Action Plan reflects setting work.53

48
 CHAPTER 3: THE POLICIES OF AI AND TRADE

Box 3.4: Case study:

Singapore’s approach to AI

For Singapore, AI is a necessary In 2019, Singapore issued a With China, Singapore is enhancing
means to overcome natural framework for responsible AI use. mutual understanding of approaches
constraints, such as a small labour The Model AI Governance to AI governance, such as under the
force on a small landmass, and to Framework provides detailed inaugural Singapore‑China Digital
raise the productivity and strengthen and practical guidance to address Policy Dialogue. Singapore also
the competitiveness of its industries, key ethical and governance issues participates in the G7 Hiroshima
both in globally tradable sectors, when deploying AI solutions. Process, the AI Safety Summit
such as trade, finance, and in In 2024, Singapore further series, the OECD AI Principles,
domestic services, such as retail extended the Model Framework the Global Partnership on AI (GPAI)
and food and beverages. beyond traditional AI to address and the World Economic Forum’s
generative AI and the novel risks it AI Governance Alliance.
For example, Singapore has poses. Within the Association of
leveraged AI in order to continue Southeast Asian Nations (ASEAN), In 2022, Singapore launched AI
to act as a global hub facilitating Singapore has spearheaded the Verify, an AI governance testing
trade and connectivity. Singapore’s development of an ASEAN Guide framework and a software toolkit,
Changi Airport, which handled more of AI Governance and Ethics. At the which contains baseline standardized
than 59 million travellers last year, United Nations, Singapore convenes tests, covering core principles
uses AI to screen and sort baggage, the Forum of Small States (FOSS), of fairness, explainability and
and to power facial recognition a grouping of 108 small economies, robustness. In 2024, Singapore
technology for seamless immigration and introduced a Digital Pillar in launched AI Verify Project Moonshot,
clearance. The Port of Singapore, 2022, which provides baseline which broadens the original toolkit
which handled cargo capacity of capacity-building for issues including to cover generative AI and return
39 million twenty-foot equivalent AI, most recently through the AI intuitive results on the quality and
unit (TEUs) in 2023, uses AI to Playbook for Small States, which safety of large language models.
direct vessel traffic, map anchorage was co-developed with Rwanda Given that the science of AI testing
patterns, coordinate just-in-time and launched at the UN Summit and governance is still nascent,
cargo delivery, process registry of the Future in September 2024. Singapore has also set up the AI
documents, and more. To facilitate Verify Foundation to harness the
communications and business Singapore works closely with a collective power and contributions
exchanges across a linguistically range of partners, bilaterally and in of the global open-source community
diverse region of 680 million people various groupings, on guidelines for to jointly develop AI Verify testing
who speak over 1,200 different AI developments and innovations. tools. The Foundation has grown
languages, Singapore has also With the United States, Singapore to more than 110 members and
invested in developing the world’s has deepened information-sharing includes companies such as
first large language model tailored and consultations on international Google, IBM, Microsoft, Red Hat,
to Southeast Asia’s languages and AI security, safety, trust and Meta and Salesforce.
cultures; this open-source model standards development through
is dubbed SEA-LION, short for collaborations in AI, including Source: Based on inputs from
Southeast Asian Languages in the US‑Singapore Critical and the Ministry of Digital Development
One Network. Emerging Technologies Dialogue. and Information, Singapore.

The heterogeneity of domestic initiatives may lead to Unintended fragmentation extends to non-AI-specific,
unintended fragmentation. Analysing eleven AI rulebooks sector-specific legislation, such as AI-relevant IP
from seven jurisdictions (i.e., Argentina, Brazil, Canada, and data regulations. Approaches to copyright “fair use”,
China, the European Union, the Republic of Korea and the for example, differ significantly across jurisdictions (see
United States), Fritz et al. (2024) find that governments Chapter 3(a)(iv)). While Japan modified its Copyright Act in
prioritize different objectives with their AI regulation, use 2018 to allow machine learning models to use copyrighted
substantially different regulatory requirements to achieve works for any purpose, including commercial use, without
the same priorities, and choose different scopes and needing explicit permission from copyright holders,58 the
formulations to achieve the same regulatory requirement EU AIA is much less permissive. According to the EU AIA,
for a shared priority, leading to unintended fragmentation at the provider of a generative AI model, whether open-source
the level of priority, requirement and scope. or closed, must establish a policy to respect EU copyright law,

49
CHAPTER 3: THE POLICIES OF AI AND TRADE

Box 3.5:
The challenge of navigating AI regulations:
The case of Canvass AI

Invited to speak at a WTO to regulate AI make achieving border market entry. She added
workshop on regulatory global reach difficult. Divergent that “minimizing complexity and
cooperation on digital products, regulations strain resources, make promoting convergence would
Humera Malik, CEO of Canvass the navigation of rules without greatly ease compliance efforts”.
AI, a startup that provides specialized knowledge difficult,
industrial AI solutions to enhance and impact market entry. Moreover, Source: [Link]
operational efficiency, profitability, data protection regulations impose english/tratop_e/tbt_e/
and sustainability, explained additional restrictions, affecting tbt_2006202310_e/
that the diverse approaches AI development and cross- tbt_2006202310_e.htm.

including the EU Directive 2019/790 on Copyright and authorities in various jurisdictions to put in place measures
Related Rights in the Digital Single Market (CDSM). to promote the development of AI. These include the
Under the CDSM, research organizations are permitted to creation of “AI factories”, to give AI start-ups and small
reproduce and extract copyrighted works for text- and data- businesses access to supercomputers on which to build
mining purposes without requiring the authorization of the their own models, research initiatives to connect researchers
copyright-owner, provided that these research organizations and educators to computational, data and training
have lawful access to the works, and that the use is for resources to advance AI research and research that employs
the purposes of scientific research. The use of copyrighted AI, and subsidies for firms that purchase domestically
materials for text- and data-mining for any reason is produced AI chips. Some of these measures appear to
also permitted beyond scientific research, but in this limit opportunities to domestic entities or to provide
context, copyright-owners have the option explicitly to reserve incentives on the condition that domestic products are used
their rights and thereby prevent the use of their works for (Aaronson, 2024b).
text- and data-mining without their approval (European
Parliament, 2024). Similarly, the AI Bill pending adoption in The economic costs of regulatory fragmentation
the Brazilian Congress provides, for example, for a limited highlight the importance of mitigating regulatory
copyright exception when the extraction, reproduction, heterogeneity. The impact of fragmentation can be felt at
storage and transformation taking place in data- and text- various levels, including lost trade opportunities, diminished
mining processes are carried out by research and journalism productivity gains and stifled innovation, with potentially
organizations and institutions, museums, archives and important economic consequences for vendors of AI-
libraries. As for the United States, while there is still enabled goods and services (Fritz and Giardini, 2024).
no legislation or regulation on this issue, a high profile As AI technologies become increasingly embedded in
ongoing litigation case was filed by the New York Times goods and services across a wide range of sectors, in the
against OpenAI for the unauthorized use of its content absence of efforts to mitigate regulatory heterogeneity,
in December 2023. Another example is the diverging the resulting costs and other negative impacts are likely to
approaches to algorithmically authored works (see Chapter grow significantly. The impact is likely to be particularly
3(a)(iv)). While the United Kingdom protects algorithmic important for small businesses, which are already struggling
creations, albeit without recognizing AI itself as an author,59 to navigate through divergent regulatory approaches on AI
Australia and the United States make it clear that a (see Box 3.5).
human author is needed (Liu and Lin, 2020). Finally,
some jurisdictions provide expansive protection to trade
secrets, applying proprietary protection to source code, Data regulations
algorithms, training materials and datasets used to train AI
models, while others do not provide them with exclusive Regulating data stands high on policy agendas. With
IP protection (Kilic, 2024). Beyond IP, data regulations the rise of digital technologies, including AI, initiatives
are also marked by a high level of fragmentation. promoting access to data to foster domestic innovation
and competition, protecting privacy and controlling
The design of some measures may affect market the flow of data across borders stand high on policy agendas.
competitors in other economies and have trade- However, what is emerging is a landscape of measures
distortive effects, leading to further fragmentation. that is not only fragmented, but that may also have trade-
The significant economic potential of AI is leading political distortive impacts beyond fragmentation.

50
 CHAPTER 3: THE POLICIES OF AI AND TRADE

Open government data and data-sharing European Data Spaces – in strategic domains, involving
initiatives to foster innovation and competition both private and public players. Training AI systems is
listed as one of the key benefits of the initiative (European
An increasing number of jurisdictions is taking Commission, 2024b). The EU Data Act, which entered into
initiatives to promote open government data to foster force in January 2024, complements the DGA and creates
business creation and innovation and to increase the processes and structures to facilitate data-sharing by
competition in domestic markets. Recognizing the value companies, individuals and the public sector (European
of data as a public good, some jurisdictions, both in Commission, 2024a). The Act protects EU businesses in
developed and developing economies,60 are pursuing open data-sharing contracts from unfair contractual terms that
government data initiatives to promote business creation may be imposed unilaterally by one contracting party
and innovation and stimulate the domestic digital and AI on another; the aim is to enable small businesses, in
economy by encouraging the use, reuse and free distribution particular, to participate more actively in the data market.
of government datasets under open data licences. Examples Other economies that have put in place data-sharing
include the EU’s Open Data Directive, India’s Open initiatives include Colombia, Japan and the Republic
Government Data platform61 and Singapore’s “Smart Nation” of Korea. Some jurisdictions, such as Australia and the
initiative.62 These initiatives come in addition to ex ante European Union, are also experimenting with legally
competition regulations put in place in some markets to better mandated data-sharing to foster a competitive environment
address competition issues raised by the digital economy.63 in which AI startups also have access to large datasets
Open government data is a goal that is also being pursued (Mayer-Schönberger and Ramge, 2018; Prüfer, 2020).
at the regional and international levels, including in the
context of the WTO Joint Statement Initiative on E-commerce The extent to which open government data and
(see Chapter 4(a)(v)). data-sharing initiatives support innovation and level
the playing field both within and across economies
Other approaches aim to promote data-sharing remains unclear. There are concerns that such initiatives
across sectors to foster innovation or to mandate it may in fact disproportionately benefit large AI firms, as
to counterbalance winner-takes-all dynamics in the these have the capacity to collect open data and to correlate
digital economy. The EU Data Governance Act (DGA), it with the “closed data” they possess and control to generate
for example, seeks to increase trust in data-sharing and data new data. As a result, large AI firms stand to gain more
availability. It entered into force in 2023 and supports the than those who lack such capabilities and have to rely on
setup of trustworthy data sharing systems – called Common open data entirely, which could amplify the growing AI divide

Figure 3.6: Data localization is growing and becoming more restrictive

(number of measures)

120

100

80

60

40

20

0
1967

1995

1997

2000

2004
1992

1993

1999

2005

2006

2007

2009

2010

2011

2012

2013

2014

2020

2021
2015

2016

2017

2022
2018

2019

Draft

Storage only Storage and flow condition Storage and flow prohibition

Source: Del Giovane et al. (2023).

51
CHAPTER 3: THE POLICIES OF AI AND TRADE

between companies (see Chapter 3(a)). Such policies with two data privacy agreements brought down by the
could also have geopolitical implications, as those operating European Court of Justice of the European Union.64
out of relatively big, closed digital economies are able to
capture open data elsewhere in addition to the data they Cross-border data flow restrictions and data
collect domestically without much external competition, which localization requirements
could result in further imbalances across economies
(Streinz, 2021). Cross-border data flow restrictions aim to limit the
flow of data, and measures to control where data is
In addition, while some-data sharing initiatives stored or processed are on the rise. Motivations behind
are clearly open to foreigners, uncertainty remains cross-border data flow restrictions and data localization
concerning other initiatives. These could raise potential requirements (i.e., explicit requirements that data be stored
most-favoured-nation (MFN) issues and result in trade- or processed domestically) vary, ranging from concerns
distortive effects. Japan, for example, announced in 2024 over sensitive data, related to national security, to privacy
that its data spaces would be open to foreigners, but the considerations. Such measures are sometimes seen as
programmes of some other jurisdictions seem designed an incentive to boost local competitiveness (Aaronson,
to support data-sharing within the jurisdiction concerned, 2024b; McKinsey, 2022). By early 2023, there were 96
which could have a trade-distortive effect (Aaronson, 2024). data localization measures across 40 economies in place,
with nearly half of the identified measures having emerged
Privacy and data protection after 2015 (see Figure 3.6). Not only has the number of
data localization measures increased, but the measures
Over the last decades, many governments have themselves are also becoming more restrictive, with more
enacted regulations for personal data protection to than two-thirds of identified measures involving not only a
address growing concerns over privacy. According to storage requirement but also a prohibition for data to flow
UN Trade and Development (UNCTAD), more than 70 per from one economy to another (Del Giovane et al., 2023).
cent of jurisdictions – 137 out of 194 – adopted legislation These data regulations apply to different types of data,
to secure the protection of data and privacy in 2021, including personal data, and to different sectors. As noted in
with significant differences across levels of development Chapter 3(a), striking the right balance between fostering AI
(UNCTAD, 2021a). The share of jurisdictions having passed innovation through access to data and protecting privacy is
such legislation is lowest in LDCs (48 per cent). The most crucial for maximizing the benefits of AI for international trade.
well known of these is the EU’s General Data Protection
Regulation, which became effective in May 2018. The global fragmentation of data flow regulations
underscores the need for increased international
AI raises new privacy concerns for individuals cooperation. While there are legitimate reasons for diversity
and consumers. This is leading to an increasingly complex in regulation, the current landscape is increasingly complex
trade-off between the need to access large amounts of data and fragmented, imposing additional costs on firms, especially
to train AI models and privacy concerns. As seen in Chapter those located in small markets, creating uncertainty, and
2, AI’s reliance on large amounts of data, including personal hindering the cross-border flow of data that plays such
data, and its capacity to process and analyse vast datasets an essential role in AI development and innovation, in particular
and to correlate data can lead to privacy breaches and for small economies. The economic costs of the fragmentation
information spillovers, introducing new privacy challenges. of data flow regimes along geo-economic blocks are
potentially sizeable, amounting to a loss of more than 1 per
Privacy and personal data protection regulations cent of real GDP, according to an OECD-WTO study (OECD
differ markedly across jurisdictions, affecting the flow and WTO, 2024). A global approach that balances the
of data. Most governments have introduced data protection need for robust data oversight and protection of privacy,
laws, but these regulations vary significantly from one while ensuring that data can be accessed and can flow
jurisdiction to another. Whereas some economies, like the freely across borders, is needed (Jones, 2023).
United States, primarily rely on the industry to self-regulate
the protection of personal data, others follow different
approaches that focus on state intervention to defend state Border measures
sovereignty, citizens’ rights, security or domestic development
(Bradford, 2023; Jones, 2023; Mitchell and Mishra, Many of the hardware components and raw materials
2018; UNCTAD, 2021b). These include limitations on the crucial to AI systems face increasing export
international transfer of personal data, aimed at maintaining restrictions. Export restrictions applied to industrial raw
jurisdictional oversight. These different approaches to data materials, many of which play a critical role in the manufacturing
governance are creating distinct “data realms” that are of advanced chips needed to power AI systems and in
fostering a new digital divide between these jurisdictions and communications equipment, increased more than five-fold
others that are rule-takers, creating regulatory uncertainty between 2009 and 2020 (OECD, 2023b). More recently,
and barriers to the flow of data across borders (Aaronson the race to dominate AI development, combined with broader
and Leblond, 2018; Jones, 2023). The divergence in economic, geopolitical and security considerations linked to
regulatory approaches between the European Union and the dual-use nature of AI systems, has led a growing number
the United States has been a particular case in point, of advanced economies to impose export restrictions on

52
 CHAPTER 3: THE POLICIES OF AI AND TRADE

advanced chips central to AI systems and on the tools used AI governance working group focused on advancing shared
to manufacture them.65 In reaction, China, one of the main principles for safe, trustworthy and responsible AI innovation,
targets of these measures, requested consultations under and calls for strengthened collaboration through joint
the WTO Dispute Settlement Understanding (DSU) in research and educational funding and for exploring reciprocal
December 202266 and imposed export restrictions on two certification programmes for American and Singaporean AI
metals used in chipmaking and communications equipment in professionals on the basis of shared standards, tests and
July 2023. benchmarks. The Dialogue includes cooperation on standard
development and a mapping exercise between domestic
There is a risk that these restrictions will affect the standard-setting bodies to align approaches. And in April
global development and deployment of AI technologies 2024, the United States and Uruguay signed a Memorandum of
and increase economic and, potentially, technical Understanding (MoU) to foster cooperation on certain critical
fragmentation. In the short term or when limited alternatives and emerging technologies, such as semiconductors, AI,
are readily available, restrictions can impact access to the data flows, telecommunications and cybersecurity, including
technology by importing economies. A longer-term effect may by identifying opportunities to support the development and
be that new technological developments will be postponed use of relevant international standards and by encouraging
due to a lack of access to advanced technology, interoperability and global compatibility, as well as greater
compounding risks of economic and technical fragmentation. cooperation in multilateral and international organizations.

China’s bilateral initiatives prioritize AI safety and

(ii) Bilateral and regional governance, as well as development issues. Dialogue
cooperation initiatives between China and the United States primarily focuses on
AI safety and governance. Announced in November 2023,
to address AI the first dialogue took place in May 2024 (The White House,
2023b; Murgia, 2024). Global governance of AI is also of
high importance in China’s discussions with African leaders
The increasing number of bilateral and regional in the context of the China-Africa Internet Development
cooperation initiatives on AI governance focusing on and Cooperation Forum. The last forum, which took place
different priorities adds to the risk of creating multiple in April 2024, called for more representation of developing
fragmented approaches. economies in the regulation of AI. Beyond the above-
mentioned examples, agreements to maintain bilateral
Bilateral cooperation initiatives that touch upon dialogues on AI have also been included in some regional
issues relevant to AI and trade prioritize different trade agreements (RTAs) and digital economy agreements
issues. Cooperation between the United States and the (see Chapter 3(b)(iii).
European Union in the context of the Trade and Technology
Council (TTC), which was established to promote EU- Approaches to AI governance initiated at the regional
US cooperation, focuses primarily on aligning terminology level take different forms. Various regional initiatives
and taxonomy and on monitoring and measuring AI risks have emerged in Africa, Asia and Latin America. Some of
(NIST, 2021). The first AI-related outcome of the TTC these take the form of ministerial declarations, such as the
was the launch, in December 2022, of a Joint Roadmap November 2023 Southern Common Market (MERCOSUR)
on Evaluation and Measurement Tools for Trustworthy AI Ministerial Declaration on the principles of human rights in
and Risk Management. The Joint Roadmap aims to guide the field of artificial intelligence,70 the October 2023 Santiago
the development of tools, methodologies and approaches Declaration to Promote Ethical Artificial Intelligence in Latin
to AI risk management and trustworthy AI, to develop a America and the Caribbean,71 and the May 2018 Declaration
common understanding of key terms, to support and lead on AI in the Nordic-Baltic Region by the Nordic Council
development of international standards, and to monitor and of Ministers.72 Others take the form of guides, such as the
measure existing and emerging AI risks. In May 2023, the TTC February 2024 Association of Southeast Asian Nations
adopted the “EU-US Terminology and Taxonomy for Artificial (ASEAN) Guide on AI Governance and Ethics (ASEAN,
Intelligence – First Edition”, which builds on existing standards 2024), or strategy documents, such as the 2024 African
such as International Organization for Standardization (ISO) Union Development Agency-New Partnership for Africa’s
standards to define key AI-related terms, but does not define Development (AUDA-NEPAD) White Paper,73 which led to
AI.67 In April 2024, the European Union and United States the adoption, on 17 June 2024, of the African Continental
launched a Research Alliance in AI for the Public Good. The Artificial Intelligence Strategy.74
Research Alliance aims to foster scientific cooperation to
better harness AI for the benefit of the environment, energy Some regional initiatives prioritize human rights and
optimization, disaster reduction and emergency responses.68 ethics, while others focus on economic development
and growth. The MERCOSUR Declaration strongly
Other bilateral initiatives in which the United States emphasizes human rights and transparency, stressing the
is involved focus more on collaboration to promote importance of avoiding discrimination, and of privacy and the
alignment in general terms. In October 2023, the integrity of information for democracy and the preservation
United States and Singapore launched a Critical and of culture, while the Santiago Declaration focuses on human
Emerging Technology Dialogue,69 which establishes a bilateral rights and ethics. The ASEAN Guide encourages alignment

53
CHAPTER 3: THE POLICIES OF AI AND TRADE

on AI governance and ethics standards based on seven in particular those signed by the United Kingdom, also
guiding principles,75 but does not list inclusive growth, recognize the importance of a risk-based and outcome-
sustainable development and well-being – as per Principle 1 based approach and of the principles of technological
of the OECD AI Principles76 – as a key principle. The ASEAN interoperability and technological neutrality,81 and include
Guide includes recommendations for both domestic and various cooperation provisions on exchanging information and
regional initiatives77 that governments in the ASEAN region sharing experiences and good practices on laws, regulations,
can take to ensure the responsible design, development, and policies, enforcement and compliance;82 ethical use, human
deployment of AI systems. Meanwhile, the AUDA-NEPAD diversity and unintended biases, industry-led technical
White Paper and the African Union Continental Artificial standards and algorithmic transparency;83 research;84 and
Intelligence Strategy focus mainly on harnessing the potential playing an active role in international fora,85 with the UK-
of AI for economic development and growth, while promoting Australia and UK-Ukraine agreements explicitly referring to
ethical use, minimizing potential risks and leveraging cooperation in the development of international standards,
opportunities. The white paper stresses the importance of regulations and conformity assessment procedures.
promoting innovation and building African multilingual tools
through AI to support a “pan-African renaissance with AI” Several AI-specific provisions explicitly refer to
and lists five pillars of action: human capital development trade. Three agreements – United Kingdom-Ukraine,86 United
for AI, infrastructure and data, enabling environments for AI Kingdom-Singapore87 and United Kingdom-Australia88 –
development and deployment, AI economy and encouraging explicitly recognize the role of AI in promoting competitiveness
investment in AI, and building sustainable partnerships. The and facilitating international trade. The United Kingdom-
Continental Strategy, adopted in June 2024, identifies four Australia agreement also encourages activities aimed at
priority sectors: agriculture, healthcare, education and climate facilitating and promoting trade in emerging technologies,
change adaptation. Likewise, the Arab AI Working Group and the agreements between the United Kingdom and
focuses primarily on cooperation to reduce the digital divide Ukraine and between the United Kingdom and Singapore
and encourage capacity-building. encourage active participation in international fora “on matters
concerning the interaction between trade and emerging
technologies”.
(iii) R
 egional trade
agreements and digital Digital trade provisions included in RTAs are also
important for AI development and use. The number of
economy agreements RTAs with digital trade provisions has been growing steadily
since the early 2000s. The first digital trade provision can be
found in the 2000 Jordan-United States Free Trade Agreement.
AI-specific provisions have started to be incorporated By the end of 2022, 116 RTAs – representing 33 per cent
into regional trade agreements (RTAs) and digital of all existing RTAs – had incorporated provisions related to
economy agreements,78 but mainly take the form of digital trade (López-González et al., 2023). These provisions
soft – i.e. non-binding – provisions. While their typically include provisions on data flows, data localization,
incorporation into these agreements is positive, such protection of personal information and access to government
provisions will not be sufficient to prevent regulatory data, which, as seen in previous sections, play an important
fragmentation. Six agreements include AI-specific role in determining access to data needed to train AI models.
provisions. These are the United Kingdom-Australia Free Provisions that ban measures mandating disclosure of source
Trade Agreement, the United Kingdom-New Zealand Free code, software and algorithms have also been included in a
Trade Agreement, and the recently signed digital economy number of trade agreements, most notably agreements led
agreements between Australia and Singapore (SADEA), by the United States. Such provisions typically aim to protect
between Chile, New Zealand and Singapore (DEPA), between technology firms from government measures requiring trade
Singapore and the United Kingdom (UKSDEA), and between secrets to be disclosed as a prerequisite for operating in
the Republic of Korea and Singapore (KSDPA), as well as a certain industries (Jones et al., 2024). Access to source code
recently signed free trade agreement between Ukraine and can, however, be important to assess the trustworthiness of AI
the United Kingdom, which has not yet come into force. systems (see Chapter 3 (a)(iii)). In addition, prohibitions on
disclosure of source code can impact technology access and
AI provisions essentially take the form of best- market competition, and limit the availability of open-source
endeavour clauses (i.e., which require parties to do software (Jones et al., 2024). Provisions on source code
everything possible to achieve the desired result). can, therefore, have a significant impact on the development
AI-specific provisions typically recognize the increasing and use of AI and on promoting AI trustworthiness. Other
importance of AI within the global economy and include provisions related to the adoption of standards and conformity
best-endeavour clauses to either “collaborate and promote”79 assessment can also play a critical role in promoting
the development of governance frameworks to promote trustworthy AI (see Chapter 3(a)(iii)), while provisions on
trusted, safe and responsible use of AI or “to develop”80 competition in the digital market are important to address the
such frameworks taking into account international guidelines, market concentration power of AI (see Chaper3(a)(i)). Finally,
with the UK-Australia and UK-New Zealand agreements provisions on customs duties on electronic transmissions
specifically referring to the 2019 OECD Principles have been important in fostering an environment conducive
(OECD, 2019a) (see Chapter 3(b)(iv)). Some agreements, to digital trade (IMF-OECD-UN-WBG-WTO, 2023).

54
 CHAPTER 3: THE POLICIES OF AI AND TRADE

The depth of digital trade provisions included in while the United Kingdom-New Zealand Free Trade
RTAs varies significantly, reflecting diverging Agreement includes binding but non-specific language.
approaches. Analysing the digital trade provisions of 12 As for disclosure of source code, agreements led by the
agreements concluded between March 2018 and January United States and digital economy agreements include
2023, Jones et al. (2024) find a high degree of heterogeneity extensive and binding protection of source code, although
between the agreements (see Figure 3.7). For example, digital economy agreements do not mention algorithms.
while most agreements contain binding obligations on the In contrast, agreements signed by New Zealand and the
free flow of data, the United Kingdom-European Union RTA Regional Comprehensive Economic Partnership (RCEP) do
does not contain any provision on non-financial data flows. not include such provisions.89
Regarding personal data protection, agreements led by
the United States consider voluntary undertakings by Few developing economies and LDCs have negotiated
private companies as sufficient to safeguard personal data, digital trade provisions. The inclusion of detailed digital
which contrasts with the European Union’s comprehensive trade provisions tends to be more common in RTAs negotiated
approach to data protection under the EU General Data by high-income and certain middle- to upper middle-income
Protection Regulation. Language on open government data economies. Only a handful of LDCs have engaged in RTAs
takes the form of best-endeavour language in agreements that contain provisions related to digital trade (IMF-OECD-
led by the United States and digital economy agreements, UN-WBG-WTO, 2023).

Figure 3.7: The depth of digital trade provisions included in RTAs varies significantly

Free flow of financial data Protection of personal information

Localization of financial data Data localization

(non-financial)

Access to Free flow of data

government data (non-financial)

Supporting Moratorium on
data innovation customs duties on
e-transmissions

Mandatory disclosure Facilitating

of source code digital inclusion

Adoption of standards & Cooperation on

conformity assessment cybersecurity matters

Governance of Al and Competition policy in

emerging technologies digital markets

CPTPP (US) 03/2018 USMCA 11/2018 JPN-US 10/2019 DEPA (SG) 06/2020 AUS-SG DEA 08/2020 SG-UK DEA 02/2022

KOR-SG 01/2023 JPN-UK 10/2020 EU-UK 12/2020 AUS-UK 12/2021 NZ-UK 02/2022 RCEP 11/2020

Source: Authors’ visualization based on Jones et al. (2024).

Notes: The value “0” (inner circle) means that the agreement does not contain a provision on a given issue. The value “1” means that
the agreement contains a provision couched in purely hortatory, or exhortatory, language. The value “2” means that the commitment
made is binding but non-specific. The value “3” means that commitments are binding and specific, with actions to be taken (or not
taken) described in clear and precise language. The value “4” means that the commitments are binding and specific, with obligations
that are more extensive in scope and very detailed. Points located between two lines correspond to the higher value in terms of
commitments, but with flexibilities. The greater the flexibilities, the closer the point is to the lower-value inner circle.

55
CHAPTER 3: THE POLICIES OF AI AND TRADE

Disciplines on trade in services in RTAs are also an rights and the ethics of AI, such as the United Nations
important channel through which governments’ trade Educational, Scientific and Cultural Organization (UNESCO)
policies and trade obligations can affect the policy Recommendation on the Ethics of AI, while others are centred
environment for AI. However, the level of commitments around safety, security, the trustworthiness of AI or its
undertaken differs significantly across economies. Services interoperability, such as the Bletchley Declaration on AI Safety.
RTAs provide significantly higher levels of market access
and national treatment commitments than under the WTO A number of initiatives also contain various common
General Agreement on Trade in Services (GATS) for elements that have an important trade and WTO angle.
different modes of supply and services sectors, including These include:
for digital and AI-related services. For example, in the context
of computer services, all WTO members from Europe, the • the recognition of the role of regulations and standards
Middle East and North America have undertaken some (including certification procedures) in governing AI and the
market access commitments on data processing services importance of interoperability between such tools;
under the GATS and/or RTAs, and most WTO members • the need to avoid regulatory fragmentation by using
have done so in Latin America and the Caribbean (88 international standards to govern AI;
per cent) and in Asia (91 per cent). However, in Africa, • the importance of an appropriate and balanced approach to
26 per cent of WTO members have market access protecting and enforcing IP rights;
commitments on data processing services, whether • the importance of privacy, personal data protection and
under the GATS or RTAs, although that proportion will data governance;
increase when the services commitments of the African • the importance of international cooperation, coordination
Continental Free Trade Area (AfCFTA) enter into force and and dialogue.
are notified to the WTO (Roy and Sauvé, forthcoming).90
Importantly, explicit references to the WTO were included
in the Final Report of the UN AI Advisory Body. The Final
(iv) International initiatives Report stresses the need for “proper orchestration” and

to address the coordination among the many international processes and

organizations producing key documents related to AI
challenges raised by AI governance, to enable a “shared normative foundation for
all AI-related efforts”, expressly referring to various WTO
agreements, such as the General Agreement on Tariffs and
Policy initiatives Trade (GATT), the General Agreement on Trade in Services
(GATS), the Agreement on Trade-Related Aspects of
The last few years have witnessed a wave of various Intellectual Property Rights (TRIPS), the Technical Barriers
initiatives related to AI. This impetus has been driven by to Trade (TBT) Agreement, the Information Technology
the realization that the inherently international nature of the Agreement (ITA) and the Trade Facilitation Agreement (TFA)
risks and benefits associated with AI require discussion, (paragraph 76 and Figure 9 of the Final Report). The Final
cooperation and solutions that are also international in Report also notes the pivotal role of international standards
nature (see Figure 3.8).91 and regulatory cooperation and recognizes the key role of the
WTO in this area (paragraph 121).94 Finally, it recommends
These initiatives involve different stakeholders and the creation of a “Global AI Data Framework” involving a
take different forms. International initiatives involve a variety of key actors, including economies and relevant
broad range of stakeholders, including governments, international organizations, including the WTO (paragraph
intergovernmental organizations, international standard-setting 170). More detail can be found in Annex 3.
bodies and businesses. Most of these initiatives take the form
of high level principles, guidance, voluntary recommendations, Several of these initiatives also address the
scientific reports, codes of conduct, or lists of policy examples, environmental impacts of AI. This is the case, for
while others take the form of international standards. In May example, for the OECD AI Principles, the G20 AI Principles,
2024, the first binding treaty on AI – the Council of Europe the New Delhi Leaders’ Declaration,95 the UNESCO AI
Framework Convention on Artificial Intelligence and human Recommendation, the G7 Guiding AI Principles and the G7
rights, democracy and the rule of law92 – was adopted. AI Code of Conduct, the Bletchley Declaration on AI Safety,
and the UN Advisory Board on AI in the United Nations.
There are elements of complementarity among such Further, the International Organization for Standardization
initiatives, and alignment on core principles, but (ISO)/International Electrotechnical Commission (IEC) Joint
different initiatives prioritize different aspects of AI Technical Committee (JTC) 1 and subcommittee 42 (SC)
governance. Some of the common themes across various (ISO/IEC JTC 1/SC 42) is currently developing an
international initiatives include, for instance, promoting safe, international standard specifically about “environmental
secure, trustworthy, “human-centric”, ethical, transparent, sustainability aspects of AI systems.” AI policy was a key
accountable and interoperable AI, and identifying and issue at the G20 Summit in Rio de Janeiro in November 2024,
mitigating AI-triggered risks through various actions, domestic with a focus on the use of AI for sustainable development; in
policies and international cooperation.93 However, in certain the G7 Trieste Ministerial Declaration adopted in 2024, G7
instances, international initiatives appear to prioritize these economies also expressed their desire to participate in the
themes differently. Some initiatives focus on issues like human G20 AI and sustainability discussions.

56
 CHAPTER 3: THE POLICIES OF AI AND TRADE

However, there is still no global alignment on AI on core principles does not guarantee alignment on how
terminology. Global agreement over key AI terminology such principles can be implemented in practice. In the
and definitions may be a particularly important trade-related absence of strong coordination, current international initiatives
element, as it may help to ensure coherence and interoperability may not be sufficient to prevent regulatory fragmentation
and to avoid fragmentation across various domestic AI at the global level. The need to improve coordination was
regulatory regimes (Meltzer, 2023). As explained in this report, acknowledged in the Final Report (2024) of the UN AI
regulatory fragmentation can itself represent an important Advisory Body (AIAB) published in September 2024 and the
trade barrier, in particular for developing economies and micro, Global Digital Compact adopted by the UN General Assembly
small and medium-sized enterprises (MSMEs). In this respect, in September 2024. The Final Report identifies three “global
the OECD AI Principles96 contain various AI definitions, of AI governance gaps” to be addressed: a “representation”
which the definitions of an “AI system”97 and an “AI system gap, a “coordination” gap and an “implementation” gap. The
lifecycle”98 are key for the implementation of any domestic AI WTO is relevant for all three, and as noted above, specific
strategy or policy and, in particular, for regulation. The Council references to the WTO are included in various places of the
of Europe Framework Convention on Artificial Intelligence and Final Report. As for the Global Digital Compact, it includes
human rights, democracy and the rule of law99 also contains a a commitment by UN members to initiate a Global Dialogue
definition of an “AI system” which is virtually identical to that in on AI governance involving governments and all relevant
the OECD Principles.100 The ISO/IEC JTC 1/SG 42, which stakeholders (paragraph 56).
is dedicated to AI standard-setting, adopted in 2022 a
document101 containing a wide range of detailed definitions
and terminology in the field of AI. It included a definition of an International initiatives
“AI system”, which shares some similarities but also includes to close the AI divide
some differences with the definition in the OECD Principles.
Finally, while the G20 AI Principles102 have more or less Increasingly, international organizations are
integrated all of the OECD Principles, they do not expressly developing courses on AI and are integrating AI
endorse the definitions, including that of an “AI System”. in their technical assistance activities, some of
Unlike OECD and ISO/IEC, the UNESCO Recommendation which have a trade component. The International
on AI Ethics does not define AI.103 Telecommunication Union (ITU), for example, offers an online
course titled “The governance of artificial intelligence” and, in
Some initiatives seem to be moving beyond general partnership with 40 other UN agencies, the ITU launched “AI
principles or guidance into implementing more for Good,” an action-oriented global platform on AI to identify
targeted or specific actions. For instance, in order to practical applications of AI to advance the UN Sustainable
foster their knowledge on existing approaches and practices, Development Goals (SDGs).109 AI for Good includes a
the G20 launched the “Examples of National Policies to year-round online programme of webinars, with an annual
Advance the G20 AI Principles”,104 and the G20 “Policy in-person AI for Good Global Summit. Other specialized
Examples on How to Enhance the Adoption of AI by MSMEs UN agencies have developed projects focused on their own
and Start-up”.105 In 2024, the G7 announced plans to advance areas of expertise. UNESCO, for example, has developed a
its 2023 Hiroshima AI process. The planned actions include Readiness Assessment Methodology to support its members
expanding outreach to partner governments to broaden in their implementation of the UNESCO Recommendation
support for the G7 AI Guiding Principles and Code of on the Ethics of AI, and is providing targeted technical
Conduct, intensifying efforts to encourage adherence to assistance in this context through projects such as its “AI
these two instruments, and intensifying cooperation across needs assessment in African countries” programme.110
multilateral forums to promote the G7 vision for advanced AI Meanwhile, the United Nations Industrial Development
systems.106 In addition, following up on the 2023 Bletchley Organization (UNIDO) has been organizing dialogues
Declaration on AI Safety, governments have agreed to on “Empowering SMEs in Developing Countries through
convey a panel of experts to produce an Intergovernmental Artificial Intelligence”111 to promote AI adoption by MSMEs
Panel on Climate Change (IPPC)-like “State of the Science” in developing economies, to enhance their competitiveness
Report,107 which will aim to review the latest cutting-edge and sustainability through shared conversations. A related
research on the risks and capabilities of frontier AI models. publication by UNIDO includes practical recommendations
The interim International Scientific Report on the Safety of and tools to help MSMEs navigate challenges and leverage
Advanced AI was published in May 2024108 and summarizes AI for various business functions and production areas.
the best of existing research, while identifying areas of research As for the World Bank, two notable projects with an AI
priority. It does not make policy or regulatory recommendations, dimension are the “Machine learning in Algeria” project,
but instead aims to inform both domestic and international which aims to enhance efficiency and integrity in customs
policymaking. The final report is expected to be published operations using machine learning, and “Fraud analytics in
ahead of the next AI summit which is expected to be held in Kenya using AI applications”, which aims to improve revenue
February 2025 in France (see also Annex 3). collection through anti-fraud measures.112 And the United
Nations Interregional Crime and Justice Research Institute
The significant overlap between initiatives, the (UNICRI) has developed a course for law enforcement
differing priorities and the lack of agreement on key agencies to equip them with the necessary resources to
terminology could create implementation challenges. institutionalize responsible AI, ensuring its alignment with
This may limit efforts to prevent fragmentation. Alignment human rights and ethics.113

57
CHAPTER 3: THE POLICIES OF AI AND TRADE

Figure 3.8: Key international policy initiatives in the area of AI

May 2019
OECD, AI Principles

June 2019
G20, AI Principles

November 2021
UNESCO, Recommendation on
the Ethics of AI

May 2023
G7, Hiroshima Process on
Generative AI

October 2023
G7, AI Guiding Principles, AI Code of
Conduct

November 2023
AI Safety Summit, “Bletchley Declaration”
on AI Safety

March 2024
UN General Assembly, AI Resolution

May 2024
International Scientific Report on the
Safety of Advanced AI (interim report)

May 2024
Council of Europe, Framework
Convention on AI, Human Rights,
Democracy and the Rule of Law

May 2024
Seoul Summit, agreement to
launch an international network
of AI Safety Institutes*

September 2024
Publication of the Final Report
of the UN AI Advisory Body

September 2024
Adoption of the UN Global
Digital Compact.

* Signatories include Australia, Canada, the European Union, France, Germany, Italy, Japan, the Republic of Korea, Singapore,
the United Kingdom and the United States.

58
 CHAPTER 3: THE POLICIES OF AI AND TRADE

resources for model training, sandboxes and curated datasets

The UN AI Advisory Body (AIAB) has called for the “to catalyse local empowerment for the SDGs”.114 The Global
establishment of a global fund for AI. Published Digital Compact,115 which was adopted by the United Nations
in September 2024, the Final Report of the UN AIAB General Assembly in September 2024 after the publication of
recommends the creation of a global fund for AI to “put the UN AIAB report, calls for “innovative voluntary financing
a floor under the AI divide”. Managed by an independent options for artificial intelligence capacity-building that take
governance structure, the fund would receive financial and into account the recommendations of the High-level Advisory
in-kind contributions from public and private sources to Body on Artificial Intelligence on a Global Fund on AI”.
facilitate access to AI enablers, such as shared computing

Endnotes
1 See [Link] stored within corporate networks, can inadvertently reveal
information about personnel involved in data collection or analysis.
2 See [Link]
In addition, metadata in online communications, such as phone
3 See [Link] numbers, emails or IP addresses, can make users identifiable
even if the data do not directly reveal personal identities
4 See [Link] (Lee-Makiyama, 2018).
5 See [Link] 12 See [Link]
6 See [Link] releases/2024/10/10/eu-brings-product-liability-rules-in-
intelligence-cs-phds-female. line-with-digital-age-and-circular-economy/#:~:text=The%20
EU’s%20product%20liability%20regime,caused%20the%20
7 At the same time, AI also holds procompetitive potential. For injury%20or%20damage and [Link]
instance, it empowers consumers to utilize abundant data for content/EN/TXT/PDF/?uri=CELEX:52022PC0496.
personalized products and transactions, and guides them in
navigating complex or uncertain markets to select the best 13 Trustworthiness is mentioned in international AI principles
offers based on preferences. This may lead to the emergence and declarations as a key attribute that an AI system should
of “algorithmic consumers”, whose decision-making is partially possess. See, for example, [Link]
automated through algorithms (Gal and Elkin-Koren, 2017). publications/ai-safety-summit-2023-the-bletchley-declaration/
the-bletchley-declaration-by-countries-attending-the-
8 Various competition enforcement cases were recently launched ai-safety-summit-1-2-november-2023, [Link]
against AI companies. For example, the US Federal Trade policy/economy/g20_summit/osaka19/pdf/documents/en/
Commission (FTC) went to court to block a proposed acquisition annex_08.pdf, Organisation for Economic Co-operation and
of Arm Ltd. by Nvidia, one of the leading producers of advanced Development (2019a) and United Nations Educational, Scientific
chips powering AI, which resulted in the latter abandoning and Cultural Organization (2021). In AI terminology,
the deal (see [Link] “trustworthiness” means the “ability to meet stakeholder …
releases/2022/02/statement-regarding-termination-nvidia- expectations in a verifiable way” (e.g., via certification against
corps-attempted-acquisition-arm-ltd). The European Commission, technical specification in a regulation or standard). More
like the UK Competition Markets Authority and the FTC, also specifically, the trustworthiness of an AI system relates to its
started looking into whether the investment of Microsoft in ability to meet various expectations, for example in terms of
OpenAI constituted a merger (European Commission, 2024a). its “reliability”, “availability”, “resilience”, “security”, “privacy”,
Cognizant of the risks that AI poses for competition, the competition “safety”, “accountability”, “transparency”, “integrity”, “authenticity”,
authorities of the European Union, the United Kingdom and the “quality” and “usability”. See ISO/IEC standard 22989:2022,
United States of America issued in July 2024 a Joint Statement on sub clause 3.5.16 (Trustworthiness - definition) and clause 5.15
Competition in Generative AI Foundation Models and AI Products (Trustworthiness - concept). See also ISO/IEC TR 24028:2020.
laying out various principles for protecting competition in the AI While the composite term “safe and trustworthy” AI is frequently
ecosystem (see [Link] used, given that “safety” is subsumed into the above definition of
joint-statement-on-competition-in-generative-ai-foundation- trustworthiness, in this report, for simplicity, we will only refer to
models-and-ai-products/joint-statement-on-competition- trustworthy AI.
in-generative-ai-foundation-models-and-ai-products).
14 For instance, certain risks may be associated with AI-enabled
9 An example of this is the fact that the Google search engine can autonomous vehicles that stem not from the physical components
outperform that of Microsoft because the former has wider access of the vehicle. Instead, the AI algorithm (and how it has been
to rarer queries. Having a variety of data and, in particular, its ability trained), may lead the vehicle to “behave” in a risky manner,
to capture more rare events are also important for making better causing not only material harms (e.g., physical injuries to the
predictions (Goldfarb and Trefler, 2018). driver, passengers or pedestrians) but also, uniquely, immaterial
10 However, advancements such as federated learning, which harms (e.g., privacy, cybersecurity, etc.). See UK Parliament
allows entities in various locations to build machine learning House of Commons’ Report on Self-Driving Vehicles (HC 519,
models collaboratively, without exchanging data (it is the algorithm 15 Sep 2023), paragraph 66 (noting studies warning that “fleets
that is transferred, not the data itself), and data trusts, a system or models of self-driving vehicles could be targeted by ‘malicious,
and legal entity that manages someone’s data on their behalf, possibly terrorist, systemic hacking’”). Regulatory solutions to
could mitigate the challenges linked to cross-border data flows such immaterial risks may also present complex ethical questions,
(Bonawitz et al., 2019; World Economic Forum, 2020a). e.g., the famous “trolley problem”, whereby an autonomous vehicle
has to “choose”, for example, between colliding with an elderly
11 AI can turn even non-personal enterprise and operational person and colliding with a mother and her young child. (e.g.,
data, such as stock inventory, into privacy risks. These data, Wells (2023); Lin (2021)).

59
CHAPTER 3: THE POLICIES OF AI AND TRADE

15 “Injury to pedestrians due to the malfunction of an autonomous The strength and advantages with AI/ML are the ability to train and
vehicle AI system would be tangible physical harm. Some harms, improve the system based on new real-world data. However, the
however, such as psychological harms, may not be as tangible system also needs to be continuously safe for patients and other
or quantifiable. Other aspects of harm that may be intangible or users, as well as comply with the applicable regulations regarding,
difficult to directly observe include bias or discrimination that may for example, validation”.
disproportionately and negatively impact particular communities
23 For instance, “[c]ustomisation makes traceability and
but be difficult to observe at the level of the individual. Violations
enforcement of product safety and cybersecurity more challenging
of the fundamental right to privacy may also be intangible, such
– many products (or properties) are changing constantly” (Lund
as the non-transparent use of an employee monitoring AI system”.
et al. 2023).
OECD Working Party on Artificial Intelligence Governance:
Stocktaking for the development of an AI incident definition, 24 The EU AI Act (2024a), for instance, seems to contain certain
document EP/AIGO(2022)11/FINAL (21 Oct 2023). provisions on this issue, as it requires that AI systems be re
certified if, after deployed, they present unforeseen “substantial
16 See, e.g., Report from the European Commission to the
modifications” (as defined in Article 3(23)), i.e., “… whenever a
European Parliament, the Council and the European
change occurs which may affect the compliance of a [Link] AI
Economic and Social Committee on “The Safety And Liability
system with this Regulation (e.g. change of operating system or
Implications of Artificial Intelligence, the Internet of Things
software architecture), or when the intended purpose of the system
and Robotics”, COM(2020) 64 final (19 February 2020),
changes, that AI system should be considered a new AI system
page 8. See [Link]
which should undergo a new conformity assessment”. However,
PDF/?uri=CELEX:52020DC0064.
“changes occurring to the algorithm and the performance of AI
17 See [Link] systems which continue to ‘learn’ after being placed on the market
p/1/u/0/w/0/d/0. or put into service (i.e., automatically adapting how functions
are carried out) should not constitute a substantial modification,
18 The Agreement on Trade-Related Aspects of Intellectual
provided that those changes have been pre-determined by
Property Rights (TRIPS Agreement) addresses inter alia the
the provider and assessed at the moment of the conformity
protection of trade secrets, including imposing certain conditions
assessment”. The EU AI Act also foresees that ex post marketing
when proprietary information (“undisclosed test and other
surveillance over AI products may: “ensure that the possible risks
data”) is accessed and used by governments for regulatory
emerging from AI systems which continue to ‘learn’ after being
purposes, albeit only in the context of “marketing approval”
placed on the market or put into service can be more efficiently
(e.g., conformity assessment procedures such as product
and timely addressed”.
certification and approval) of pharmaceuticals and agricultural
chemical products (Article 39.3). Similarly, the Technical Barriers 25 ISO/IEC TR 24368 (2022) gives examples of areas in which
to Trade (TBT) Agreement requires that WTO members ensure there is an “increasing risk for undesirable ethical and societal
that the confidentiality of information in the context of conformity outcomes and harms”, e.g.,: “financial”; “psychological”; “physical
assessment procedures (e.g. product certification and approval) is health or safety”; “intangible property (for example, IP theft,
(i) respected for imported and national products “in the same way” damage to a company’s reputation)”; “social or political systems
and (ii) respected “in such a manner that legitimate commercial (for example, election interference, loss of trust in authorities)”;
interests are protected.” (Article 5.2.4). and “civil liberties (for example, unjustified imprisonment or other
punishment, censorship, privacy breaches)”.
19 Mitchell et al. (2023) contains a detailed analysis of
circumstances when regulating AI can be performed without 26 Commenting on the fact that the AI Act’s implementation
need to access source code (“white box” testing for low-risk AI may involve the adoption of technical standards for addressing
systems), and of circumstances when a deeper understanding and both material (e.g., health) and immaterial risks (e.g., fundamental
explanation of the AI system’s decisions and recommendations is rights), Smuha and Yeung (2024), observe that: “… unlike risks to
needed (high risk AI systems) and justifies requiring access to the safety generated by chemicals, machinery or industrial waste, all
code (“black box” testing). of which can be materially observed and measured, fundamental
rights are, in effect, political constructs. These rights are
20 “Evolution” in the sense that some AI systems allow the
accorded special legal protection so that an evaluation of alleged
product to better perform, adapt and finetune overtime for a
interference requires close attention to the nature and scope of
given circumstance or for a given user, as it works in practice
the relevant right and the specific, localized context in which a
and receives and crunches more data; a sort of “personalized
particular right is allegedly infringed. We therefore seriously
AI product” similar to the idea of “personalized medicine” (e.g.,
doubt whether fundamental rights can ever be translated into
using knowledge of a patient’s genetic profile to select “the proper
generalized technical standards that can be precisely measured
medication or therapy and administer it using the proper dose or
in quantitative terms, and in a manner that faithfully reflects
regimen” – see [Link]
what they are, and how they have been interpreted under the
Personalized-Medicine). In fact, AI can be a driver and enabler
European Charter on Fundamental Rights and the European
for advancing personalized medicine in the area of genomics
Convention on Human Rights”.
medicine. See Cesario et al. (2023); World Health Organization
(2021). 27 See also WTO official document number G/TBT/GEN/356,
available at [Link]
21 In this respect, the EU AI Act (2024a), for instance, notes in its
preamble (recital 5), that “AI may generate risks and cause harm 28 See, for example, the European Union’s Artificial Intelligence
to public interests and fundamental rights” and that “[s]uch harm Act (AIA) ([Link] under which all
might be material or immaterial, including physical, psychological, providers of “general-purpose AI models” are mandated to create
societal or economic harm.” (italics added). technical documentation that details the training and testing
processes, establish a copyright policy, and provide a sufficiently
22 Lund et al. (2023) present a useful example of this tension
detailed summary of the content used for training. In contrast,
with respect to medical devices, explaining that one of the “main
free and open AI models are only obliged to establish a copyright
obstacles of using AI in healthcare, and therefore AI-based medical
policy and submit a summary of training content.
software”, is that “how to address continuous change i.e., locked
algorithms vs non-locked autonomous systems is a challenge.

60
 CHAPTER 3: THE POLICIES OF AI AND TRADE

29 RDC nº 657/2022 – Anvisa [Link] organization, regardless of size, type and nature, that provides or
documents/10181/5141677/RDC_657_2022_.pdf/f1c32f0e- uses products or services that utilize AI systems”.
21c7-415b-8b5d-06f4c539bbc3.
41 However, some experts argue that some important provisions
30 Such as Israel, Japan, the Republic of Korea, Singapore, the of the AI Act do not follow a purely risk based approach
United Kingdom and the United States. The Global AI Index 2024 (Ebers, 2024).
published by Tortoise Media, which ranks economies by their
42 See [Link]
AI capacity at the international level: [Link]
com/intelligence/global-ai/. 43 For example, the UK has committed £100 million toward
building a “public foundation model” to support academic, small
31 While copyright protects creation of the (human) mind, patent
business and public sector applications (see [Link]
protection is available for technical innovations (by humans).
uk/government/news/initial-100-million-for-expert-taskforce-
32 US Copyright Registration Guidance: Works Containing to-help-uk-build-and-adopt-next-generation-of-safe-ai), and the
Material Generated by Artificial Intelligence, available at https:// US National Artificial Intelligence Research Resource is working in
[Link]/ai/ai_policy_guidance.pdf; US Court of Appeals a similar direction (see [Link]
for the 9th Circuit, Naruto v. Slater, [Link] intelligence/nairr).
gov/datastore/opinions/2018/04/23/[Link]; US District
44 Standards are one of the three types of technical barriers
Court for the District of Columbia, Thaler v. Perlmutter, [Link]
to trade (TBT) measures that establish product specifications.
[Link]/cgi-bin/show_public_doc?2022cv1564-24.
They differ from technical regulations, however, as standards
33 CJEU, Infopaq International A/S v Danske Dagblades are voluntary documents. It is also not uncommon for standards
Forening, Case C-5/08 (Intellectual Property Repository, 2023; adopted by governments to be made mandatory later on, thus
Zhou, 2019). becoming technical regulations. Standards can be developed
by different entities within WTO Members, including both
34 Strategies articulate the government’s vision regarding the
governmental and non-governmental bodies (WTO, 2021).
contribution of science, technology and innovation (STI) to the
social and economic development of an economy. They set 45 See, e.g., Kerry (2024). For instance, in 2024, China
priorities for public investment in STI and identify the focus of issued draft Guidelines for AI Standardisation which proposes
government reforms, for instance in areas such as funding public to form more than 50 national and industry-wide standards
research and promoting business innovation (OECD, 2016b). and more than 20 international standards for AI by 2026 (see
[Link]
35 See OECD AI database. [Link]
guidelines-standardising-ai-industry-2024-01-17/ and https://
overview
[Link]/miit-ai/). The European Commission mandated
36 Although not a country, the European Union has the power to European standardisation organizations to develop AI-related
adopt EU-wide trade-related legislation within the parameters set standards taking into account that standards will play an
by its founding treaties. important role in fulfilling requirements under the EU AI Act
([Link]
37 See [Link] ref=C(2023)3215&lang=en).
ip_24_383.
46 For example, China’s standards strategy of 2021 identified AI
38 See [Link] as one of the key areas. See Kerry (2024).
39 Regulation (EU) 2024/1689 of the European Parliament and 47 This is based on the data from the AI Standards Hub and is
of the Council of 13 June 2024 laying down harmonised rules provided for illustration purposes only. This data is presented
on artificial intelligence and amending regulations (EC) No without prejudice, and should not be understood as a position,
300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) on whether these documents are “standards” within the definition
2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives of Annex 1 of the TBT Agreement. See AI Standards Search -
2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial AI Standards Hub.
Intelligence Act). AIA was published in the EU Official Journal
on 12 July 2024 and entered into force 20 days later. However, 48 See the AI Standards Hub at [Link]
most of AIA’s rules are only applicable 24 months after its entry
49 For example, the WTO Technical Barriers to Trade (TBT)
into force, although it provides for shorter applicability periods
Agreement requires that WTO members use relevant international
with respect to certain rules (e.g., bans on “prohibited practices”
standards as a basis of their domestic standards, technical
that are listed as posing “unacceptable risks” will already apply six
regulations and certification procedures (see Chapter 4).
months after entry into force), as well as longer periods for others
(e.g., 36 months for certain “high risk systems” covered by existing 50 See [Link]
EU harmonization legislation and for general-purpose AI systems h t t p s : / / w w w. i n d u s t r y. g o v. a u / d a t a - a n d - p u b l i c a t i o n s /
on the EU market before the Act applies to them). australias-artificial-intelligence-action-plan.
40 In the notification in 2021 of a draft of the AIA (document G/ 51 See [Link]
TBT/N/EU/850), the European Union explained that this proposal t20231024_11167412.htm.
was meant to provide: “… a set of recommendations intended
52 See [Link]
to help the organization develop, provide, or use AI systems
actions/2023/10/30/executive-order-on-the-safe-secure-
responsibly in pursuing its objectives and meet applicable
and-trustworthy-development-and-use-of-artificial-intelligence/.
requirements, obligations related to interested parties and
See also NIST “A Plan for Global Engagement on AI Standards”
expectations from them. It includes the following: approaches to
(final, July 2024 - available at: [Link]
establish trust in AI systems through transparency, explainability,
[Link]) and US Government National Standards
controllability, etc.; engineering pitfalls and typical associated
Strategy for Critical and Emerging Technology presented at the
threats and risks to AI systems, along with possible mitigation
TBT Committee meeting held on 21-23 June 2023 ([Link]
techniques and methods; and approaches to assess and
[Link]/dol2fe/Pages/SS/[Link]?filename=q:/G/
achieve availability, resiliency, reliability, accuracy, safety, security
TBT/[Link]&Open=True, paragraph 6.32).
and privacy of AI systems. This document is applicable to any

61
CHAPTER 3: THE POLICIES OF AI AND TRADE

53 For example, Australia has been actively engaged in the work Digital%20Markets%2C%20Competition%20and,in%20
of the International Organization for Standardization (ISO) and and%20innovate%20new%20technology.
International Electrotechnical Commission (IEC) Joint Technical
64 The Safe Harbour Privacy Principles, which were developed
Committee (ISO/IEC JTC1/SC42) and, in 2024, Australia
between 1998 and 2000 to prevent private organizations within
announced the adoption of one of the ISO/IEC JTC1/SC42
the European Union or United States that store customer data
standards (see [Link]
from accidentally disclosing or losing personal information, were
australia-adopts-the-international-standard-for-ai-management-
brought down by the European Court of Justice (ECJ) in 2020
system-as-iso-iec-42001-2023).
after Max Schrems, an Austrian activist, lawyer and author brought
54 EU AI Act, Preamble, Recital (176). a case against Facebook for its privacy violations, including
violations of European privacy laws and the alleged transfer
55 While there is no globally agreed definition, the European
of personal data to the US National Security Agency (NSA) as
Parliament Research Service notes in its paper on “Artificial
part of the NSA’s PRISM data-mining programme. The Safe
Intelligence Act and Regulatory Sandboxes” ([Link]
Harbour Privacy Principles were replaced with the Privacy Shield
[Link]/RegData/etudes/BRIE/2022/733544/
until 2020, when the ECJ once again brought it down. A new
EPRS_BRI(2022)733544_EN.pdf) that “regulatory sandboxes
agreement was reached in July 2023 to allow data flows based
generally refer to regulatory tools allowing businesses to test
on the “adequacy decision” mechanism of the EU General Data
and experiment with new and innovative products, services or
Protection Regulation.
businesses under supervision of a regulator for a limited period
of time. As such, regulatory sandboxes have a double role: 1) 65 The United States initiated export controls on semi-conductors
they foster business learning, i.e. the development and testing of in 2022, and these restrictions were broadened over time.
innovations in a real-world environment; and 2) support regulatory In 2023, the Netherlands imposed restrictions on high end
learning, i.e. the formulation of experimental legal regimes to chipmaking. The United Kingdom, Canada and Japan followed
guide and support businesses in their innovation activities under with their own restrictions. See Financial Times (2022) and
the supervision of a regulatory authority”. Wolff (2022).
56 See [Link] 66 See [Link]
ds615_e.htm.
57 i.e., the initial version of a product that includes only the core
features necessary to meet basic user needs and gather feedback 67 See [Link]
for future improvements. terminology-and-taxonomy-artificial-intelligence.
58 WTO official documents IP/N/1/JPN/36, IP/N/1/JPN/C/6 and 68 See [Link]
IP/C/M/92/Add.1, available at [Link] good-eu-us-research-alliance-ai-public-good.
59 The UK Copyright, Designs and Patents Act 1998 provides 69 See [Link]
that authorship is attributed to “the person by whom the releases/2023/10/12/u-s-singapore-critical-and-emerging-
arrangements necessary for the creation of the work are technology-dialogue-joint-vision-statement/.
undertaken”, paragraph 9(3). Other common law jurisdictions
70 See [Link]
such as India (copyright Act 1957 paragraph 2(d)), Ireland
2024/04/ D E C LARAC I O N-S O B R E-LO S-P R I N C I P I O S-
(Copyright and Related Rights Act 2000 21), New Zealand
D E-D E R E C H O S-H U MAN O S-E N-E L-AM B ITO-D E-LA-
(Copyright Act 1994 5(1)) and South Africa (Copyright Act 1978
[Link].
1(iv)) follow the UK approach.
71 See [Link]
60 For more information on developing economies pursuing
402a35a0-1222-4dab-b090-5c81bbf34237/declaracion_de_
open government data policies see Verhulst and Young (2017).
[Link].
61 See [Link]
72 See [Link]
62 See [Link] The OECD recently ForsAetisraduneytid/ Framtidarnefnd/AI%20in%20the%20
launched an Open Government Data project to map practices Nordic-Baltic%[Link].
across economies and assess the impact of open government
73 See [Link]
data (OECD, 2019b).
regulation-and-responsible-adoption-of-ai-in-africa-towards-
63 Unlike anti-trust policies, ex ante regulations apply at an achievement-of-au-agenda-2063.
industry or sectoral level and attempt to define how the largest
74 See [Link]
companies must compete in the market. One such set of
ministers-adopt-landmark-continental-artificial-intelligence-
regulations is the European Union Digital Markets Act (DMA),
strategy#:~:text=The%20Continental%20AI%20Strategy
which entered into force in November 2022 and became
%20provides,potential%20risks%2C%20and%20
applicable, for the most part, on 2 May 2023. The DMA is designed
leveraging%20opportunities.
to address the market power of major digital platforms, referred to
as “gatekeepers”. It aims to ensure fair competition and innovation 75 The seven guiding principles are transparency and
in the digital market by preventing gatekeepers from imposing explainability, fairness and equity, security and safety, robustness
unfair conditions on businesses and consumers (European and reliability, human-centricity, privacy and data governance, and
Commission, 2022). The DMA includes specific obligations for accountability and integrity.
these gatekeepers, such as allowing third parties to interoperate
76 See [Link]
with their services and prohibiting them from favouring their own
html.
services. The UK Digital Markets, Competition and Consumers
Bill is another example of new ex ante approach to digital markets. 77 National recommendations include nurturing AI talent and
The Bill encourages the most powerful firms in dynamic digital upskilling the workforce, supporting the AI innovation
markets to work with regulators to ensure that competition ecosystem and promoting investment in AI start-ups, investing
is maintained on an ongoing basis. See [Link] in AI research and development, promoting adoption of useful
government/news/changes-to-digital-markets-bill-introduced- tools by businesses to implement the ASEAN Guide on AI
to-ensure-fairer-competition-in-tech-industry#:~:text=The%20 Governance and Ethics, and raising awareness among citizens on

62
 CHAPTER 3: THE POLICIES OF AI AND TRADE

the effects of AI in society. The regional recommendations are: 94 i.e., calling for a “AI Standards Summit” involving key internation
to establish an ASEAN Working Group on AI Governance standard-setting bodies (e.g., International Telecommunication
consisting of representatives from member states to drive and Union (ITU), the International Organization for Standardization (ISO)
oversee AI governance initiatives in the region; to adapt the / International Electrotechnical Commission (IEC) and the Institute
AI Guide to address the governance of generative AI; and of Electrical and Electronics Engineers (IEEE)) and expressly
to compile a compendium of use cases demonstrating indicating that the WTO should be involved in these discussions.
practical implementation of the AI Guide by organizations
operating in ASEAN. 95 See [Link]
[Link].
78 Digital economy agreements are a new type of agreement.
They aim to regulate digital trade, data flows and emerging 96 See [Link]
technologies like AI. Digital economy agreements reflect html.
governments’ response to the need for regulatory frameworks 97 See the OECD revised definition of “AI system”: an “AI
tailored to the complexities of digital trade and the digital economy. system” is “a machine-based system that, for explicit or implicit
To date, four digital economy agreements have been signed and objectives, infers, from the input it receives, how to generate
have entered into force: the Singapore-Australia Digital Economy outputs such as predictions, content, recommendations, or
Agreement (SADEA), signed in 2020; the Digital Economy decisions that can influence physical or virtual environments.
Partnership Agreement (DEPA) between Chile, New Zealand Different AI systems vary in their levels of autonomy and
and Singapore, signed in 2020; the United Kingdom-Singapore adaptiveness after deployment”.
Digital Economy Agreement (UKSDEA), signed in 2022; and
the Republic of Korea-Singapore Digital Partnership Agreement 98 An “AI system lifecycle” involves the: “i) ‘design, data and
(KSDPA), signed in 2022. Others under negotiation include models’; which is a context dependent sequence encompassing
the ASEAN Digital Economy Framework Agreement (DEFA) and planning and design, data collection and processing, as well as
the EFTA-Singapore Digital Economy Agreement. model building; ii) ‘verification and validation’; iii) ‘deployment’;
and iv) ‘operation and monitoring’. These phases often take
79 KSDPA, DEPA and United Kingdom-Australia. place in an iterative manner and are not necessarily sequential.
80 United Kingdom-Australia, United Kingdom-New Zealand The decision to retire an AI system from operation may occur at any
and United Kingdom-Singapore (the latter specifies point during the operation and monitoring phase” – see OECD AI
“where appropriate”). Principles (2019), section 1.I.

81 United Kingdom-Ukraine, United Kingdom-Singapore and 99 See [Link]

United Kingdom-New Zealand. framework-convention-on-artificial-intelligence.

82 United Kingdom-Ukraine, KSDPA, United Kingdom-Singapore 100 So far, some domestic AI regulations, such as the EU’s AIA
and United Kingdom-New Zealand. and Brazil’s draft Senate Bill n. 2338/2023, seem to have adopted,
almost verbatim, the OECD Principles definitions, including that
83 United Kingdom-Ukraine, United Kingdom-Singapore and of “AI system”.
United Kingdom-New Zealand.
101 See ISO/IEC 22989:2022 (available at [Link]
84 United Kingdom-Ukraine, United Kingdom-Singapore, SADEA [Link]/ittf/PubliclyAvailableStandards/[Link]).
and United Kingdom-Australia.
102 See [Link]
85 United Kingdom-Ukraine and United Kingdom-New Zealand. [Link].
86 Article 132-V. 103 It states in this respect that it “does not have the ambition to
87 Article 8.61-R. provide one single definition of AI, since such a definition would
need to change over time, in accordance with technological
88 Article 20.4. developments. Rather, its ambition is to address those features
89 New Zealand decided to exclude provisions on source of AI systems that are of central ethical relevance”. Yet, the
code from its agreements following a November 2021 decision Recommendation does provide a broad understanding of what “AI
of the Waitangi Tribunal, which found the source code provision systems” mean, i.e., “systems which have the capacity to process
in the Comprehensive and Progressive Agreement for data and information in a way that resembles intelligent behaviour,
Trans-Pacific Partnership (CPTPP) to be in breach of the Treaty and typically includes aspects of reasoning, learning, perception,
of Waitangi after Māori tech experts argued that there was a risk prediction, planning or control” (paragraph 2). For UNESCO, such
of biased assumptions in algorithmic design or training data. broad understanding is “crucial as the rapid pace of technological
See Jones (2024). change would quickly render any fixed, narrow definition outdated,
and make future-proof policies infeasible” (UNESCO, 2023).
90 The analysis in Roy and Sauvé (forthcoming) is based on 142
RTAs notified under GATS Article V. 104 See [Link]
html.
91 See, for example, the Bletchley Declaration (2023b),
which states that, “[m]any risks arising from AI are inherently 105 See [Link]
international in nature, and so are best addressed through Annex1_DECLARATION-OF-G20-DIGITAL-MINISTERS-2021_
international cooperation”. [Link].

92 See [Link] 106 See [Link]

framework-convention-on-artificial-intelligence. page/field_ref_resources/390de76d-d4f5-4f45-a7b4-
f6879c30c389/0fbffe8a/20231201_en_news_g7_result_
93 While, so far, AI global governance initiatives have, on the [Link].
one hand, “yielded similarities in language, such as the importance
of fairness, accountability, and transparency”, on the other 107 See [Link]
hand, approaches on defining AI are less coordinated and summit-2023-chairs-statement-state-of-the-science-2-november/
coherent ([Link] state-of-the-science-report-to-understand-capabilities-and-risks-
body_interim_report.pdf). of-frontier-ai-statement-by-the-chair-2-november-2023.

63
CHAPTER 3: THE POLICIES OF AI AND TRADE

108 See [Link]

66f5311f080bdf716392e922/international_scientific_report_
on_the_safety_of_advanced_ai_interim_report.pdf.
109 See [Link]
110 See [Link]
pf0000375322/PDF/[Link].
111 See [Link]
unido/.
112 See [Link]
wbg/.
113 See [Link]
actions/.
114 See [Link]
for_humanity_final_report_en.pdf.
115 See [Link]

64
4 What role
for the WTO?
65
 CHAPTER 4: WHAT ROLE FOR THE WTO?

The WTO has an important role to play in AI avoid regulatory fragmentation. WTO rules and
governance. As seen in Chapters 2 and 3, AI can have processes promote global convergence through
a significant impact on trade and can open up many transparency, discussion and exchange of good practices,
opportunities, but it also creates various trade-related regulatory harmonization and non-mandatory policy guidance,
policy challenges. An increasing number of initiatives has as well as through the negotiation and implementation
emerged at the domestic, bilateral, regional and international of new trade rules.
levels to address risks associated with AI and to harness
its benefits, but these are creating a fragmented policy
landscape. The WTO, as the only rules-based global body (i) Promoting transparency
dealing with trade policy, can play an important role in
supporting governments to foster the growth of AI. In this
respect, WTO rules may be crucial in facilitating trade in Transparency, a key function of the WTO, provides
AI-related goods and services, promoting global WTO members with the opportunity to be kept
convergence, fostering access to and innovation in AI, abreast of the latest regulatory developments. All
avoiding discrimination, minimizing international negative WTO agreements integrate transparency provisions,
spillovers, helping to address and prevent trade tensions, including in some cases requirements that WTO members
and building capacity in AI. However, the rise of AI could also publish and promptly notify new, or any changes to
have implications for international trade rules. existing, laws, regulations or administrative guidelines that
significantly affect trade in the areas covered by WTO
agreements. Members are also required to establish
enquiry points responsible for responding to questions that
(a) Promoting global stakeholders from any WTO member may have on rules
and regulations related, for example, to services, TBT or
convergence intellectual property (IP), all of which play an important role
in AI governance.

Addressing the challenges raised by AI requires The transparency mechanism of the TBT Agreement3
global coordination and cooperation to promote goes further in promoting global convergence
regulatory convergence. If widely different, or even and coherence by requiring that members notify
conflicting, domestic regulatory approaches on AI are regulatory measures at a draft stage to the TBT
developed, unnecessary regulatory fragmentation may Committee.4 Early notifications can help governments and
ensue, and this could hamper opportunities and benefits other stakeholders to be kept abreast of proposed AI-
associated with AI and undermine public trust in this related regulations more quickly, and give members the
transformative technology. As seen in Chapter 3, discussions opportunity to voice questions and concerns regarding
on the global governance of AI have accelerated significantly upcoming regulatory measures in a timely manner. It
over the past few years. However, the different approaches also helps to ensure that comments can be taken into
are raising growing concerns about regulatory account well before measures are finalized, which can lead
fragmentation and its potentially damaging impact on to better quality regulations and lower trade costs, and
cross-border economic activities. For example, the risk it fosters understanding of members’ regulatory approaches
of regulatory fragmentation dominated discussions at and promotes more effective and globally coordinated,
the OECD Global Forum on Trade on 3 October 2023. coherent regulatory outcomes. For example, in 2021, a
Similarly, WTO members recently expressed concern developed member notified a proposal for AI regulation to
with regulatory fragmentation in this area, which they the TBT Committee (the EU AI Act)5, which was later also
considered could block opportunities and benefits discussed in the Committee in the context of a “specific
associated with such novel products, as well as undermining trade concern” (STC).6 In April 2024, for the first time,
public trust and leading to an enlargement of the digital a developing member notified an AI-specific regulation,
divide. Among other issues, they stressed the role of (“KS 3007:2024 Information technology – Artificial
closer international cooperation in building inclusive Intelligence – Code of Practice for AI Application”),
global digital governance.1 As governments recognized in to the Committee.7 More broadly, the TBT Committee
the 2023 Bletchley Declaration: “[m]any risks arising from has been receiving an increasing number of notifications
AI are inherently international in nature, and so are best of a wide range of digital-technology-related regulatory
addressed through international cooperation”.2 This was measures, including concerning the Internet of Things,
echoed again in the recent 2024 Final Report of the UN 5G, 3D printing, drones and autonomous vehicles.8
AI Advisory Body. Indeed, when it comes to trade, regulatory Transparency may also help members to “emulate more
cooperation at a global level can help build trust, avoid efficient regulatory examples” made widely available in WTO
unproductive trade frictions, and prevent unnecessary notifications (Mavroidis, 2016).
negative trade impacts without compromising legitimate
public policy objectives (OECD and WTO, 2019). An important transparency tool is the ePing SPS and
TBT Platform.9 This publicly and freely available tool
The rise of AI increases the importance of the WTO, includes an email alert service on notifications covering
and its transparency and deliberative functions, as a products and markets of interest, including AI-related
forum for cooperation and regulatory alignment to notifications. All interested stakeholders, including

66
 CHAPTER 4: WHAT ROLE FOR THE WTO?

businesses of any size, can register on the platform and track currently ongoing “Tenth Triennial Review of the operation
regulatory developments about products and markets of and implementation of the TBT Agreement”, proposals have
interest to them, and communicate with other stakeholders. been made to discuss AI specifically, or at least certain
AI-related issues, in the TBT Committee.14
The WTO Trade Policy Review Mechanism (TPRM)
also contributes to enhancing the transparency Since 1998, multilateral discussions under the WTO
of members’ trade policies. All WTO members are Work Programme on e-commerce have considered
subject to periodic reviews of their domestic trade policies. how WTO rules apply to e-commerce. These discussions
The TPRM aims to improve members’ adherence to WTO rules, intensified following the Ministerial Decision on the
disciplines and commitments, through greater transparency E-commerce Moratorium and Work Programme,15 which
in, and understanding of, WTO trade policies and practices.10 was adopted at the 12th Ministerial Conference (MC) in
In fact, the subject of AI has been raised in the context 2022 and provides a platform for experience-sharing and
of various recently concluded trade policy reviews (TPRs).11 mutual learning. Issues relevant to AI discussed under
the work programme include consumer protection, legal
and regulatory frameworks, and digital industrialization.
(ii) P
 romoting dialogue Discussions also covered the important issue of the

and exchange of digital divide.

good practices Experience-sharing on AI is also slowly emerging

in other WTO bodies. For instance, the Committee on
Sanitary and Phytosanitary Measures (SPS Committee)
The WTO provides a global forum for constructive recently held a thematic session to explore the utilization of
discussions, exchange of good practices and technological solutions, including AI and machine learning,
cooperation. In this context, governments can discuss how in the field of SPS.16 In addition, Australia recently submitted
best to design nuanced, flexible and adaptable regulatory a proposal that the future agenda of discussions and
solutions to address the goods, services and IP-related experience-sharing of the SPS Committee put a “strong
aspects of AI in a coordinated manner. Global alignment focus” on the potential application of AI technologies in
starts with dialogue, and WTO bodies provide fora to which regulatory frameworks that govern agri-food trade.17
members can bring trade-related issues they wish to explore
and discuss. Given AI’s fast-changing and complex nature, The WTO can provide a platform for governments to
nurturing dialogue and an exchange of good practices on an brainstorm on how best to design nuanced, flexible,
open, inclusive and ongoing basis is critical. coordinated regulatory solutions to address the
trade-related aspects of AI. Issues flagged by scholars
Various WTO bodies have organized thematic that could be discussed include: how to ensure that possible
discussions on AI trade-related topics to exchange regulatory solutions do not become obsolete as AI rapidly
experiences and identify good practices. Among them evolves; how to ensure a lifecycle compliance of AI and
are the Council for Trade-Related Aspects of Intellectual AI-embedded products with relevant requirements under
Property Rights (TRIPS Council) and the TBT Committee. standards and technical regulations; how to ensure post-
For instance, in 2023, South Africa called for a revitalization market surveillance of AI and AI-enabled products; and
of discussions on e-commerce-related IP matters in the how to improve the WTO’s engagement with other relevant
TRIPS Council and proposed a structured dialogue based bodies and organizations that are currently discussing and
on specific questions, including what measures members elaborating policies, guidance and international standards
are adopting to improve access to AI technologies.12 relevant for AI regulation and global governance.18 Such
In the same year, a group of “Friends of IP and Innovation”, discussions would help members to become aware of
including Australia, Canada, the European Union, Japan, each other’s different systems and to understand better the
Singapore, Switzerland, Chinese Taipei and the United similarities and divergences in their regulatory approaches.
States, proposed that, due to the immense benefits This, in turn, could provide a solid basis for further considering,
of cross-border cooperation among IP offices and the in a multilateral setting, how to ensure better regulatory
unclear application of existing IP systems to advanced coherence in the area of AI. A notable example of this
technologies, such as AI and the metaverse, it would be useful positive role of the WTO is the recently adopted 2024 TBT
for IP offices to engage in global discussions on suitable IP “Guidelines on Conformity Assessment Procedures” (CAP
protection in these technology fields, and to share domestic Guidelines) (WTO, 2024b). The CAP Guidelines not only
experiences and best practices. The TBT Committee, on its recognize the importance of digital technologies to improve
side, recently held five thematic sessions on digital issues the way governments certify products in terms of safety
and related regulatory measures with the aim of improving and quality, but also stress the importance of ensuring
global regulatory cooperation between members in these “flexibility and agility in the face of uncertainty”, including due to
areas. The thematic sessions covered intangible digital “rapidly changing technological, societal, geopolitical and
products (including AI), cybersecurity, conformity economic trends”, by ensuring conformity assessment
assessment issues with respect to products sold via procedures are “adaptive, responsive, and remain relevant”.
e-commerce, digital solutions for performing conformity
assessment, and the use of digital technologies and tools WTO committees also serve as fora for information-
in members’ regulatory processes.13 In addition, under the sharing and discussions between WTO members

67
 CHAPTER 4: WHAT ROLE FOR THE WTO?

and standard-setting organizations. Standard- While the WTO does not itself develop international
setting organizations have observer status in various WTO standards, some of its agreements explicitly
committees, including the TBT and SPS Committees. WTO encourage their use. The TBT Agreement is a particular
committees can therefore provide a valuable opportunity case in point, as it encourages members to engage in
for constructive dialogue between members and standard- regulatory harmonization by requiring them to use relevant
setting organizations to identify needs and gaps in standards international standards as a basis of their domestic
development from an international trade perspective. For standards, technical regulations and certification procedures.
example, in the June 2024 TBT Committee meeting, the ISO This requirement is strengthened by a presumption that
noted that, together with the International Electrotechnical a regulation does not create an unnecessary obstacle
Commission (IEC), it had published the joint international to international trade – which must be avoided – if it is
standard ISO/IEC 42001, which it claimed to be “the world’s prepared “in accordance with” such standards. At the
first AI management system standard”, laying down “the same time, the TBT Agreement recognizes that there may
foundation for ethical, safe, and innovative use of AI across its be legitimate reasons for an international standard not
many applications and promoted trust by effectively managing to be used as a basis for a given regulation. Members,
AI-related risks.”19 At that same meeting, the United Nations in particular developing-economy members, are thus allowed
Economic Commission for Europe (UNECE) informed to deviate from these standards under certain conditions.22
members about the work being undertaken by its Working
Party on Regulatory Cooperation and Standardization Policies To harmonize technical regulations on as wide a basis
on adopting relevant guidance on “technical regulations of as possible, the TBT Agreement strongly encourages
products/services with embedded artificial intelligence”.20 In members to “take a full part” in the elaboration
addition, during a recent Thematic Session held by the SPS and development of international standards.23
Committee, relevant work on the use of digital technologies, Active participation in international standard-setting work
including AI, was presented by various international increases the chances that a member will be a standard-
standard-setting bodies including the World Health maker rather than merely a standard-taker. This can make
Organization (WHO), the Food and Agriculture Organization international standards more inclusive, legitimate and useful
of the United Nations (FAO) Codex Alimentarius, the World as benchmarks for the promotion of regulatory harmonization
Organization for Animal Health (WOAH) and the International and coherence, including in AI regulation and standardization.
Plant Protection Convention (IPPC).21 However, it should be noted that active engagement in the
development of numerous – and usually simultaneous –
international standards could be particularly problematic
(iii) P
 romoting regulatory for developing-economy members in light of their scarce
resources and lack of relevant expertise; this is especially
harmonization and the case when the standardization process involves new
coherence through technological fields that are complex and fast evolving.
In this context, the TBT Agreement requires members to
international standards, advise developing-economy members, upon request, and
mutual recognition to grant them technical assistance regarding participation in
international standardizing bodies24 (see also Chapter 4(e)).
and equivalence
However, certain aspects of international
standardization in the area of AI may be challenging.
International standards play an important role in Indeed, it might be difficult, or, to some, even inappropriate
promoting global regulatory alignment and coherence. (Pouget, 2023), to agree on a common international
The development and use of international standards in denominator with respect to certain AI-related societal values
the area of AI can provide a common benchmark when and concerns such as ethical or moral values, the relative
governments design and adopt standards or regulations on importance of which may vary across economies and
AI systems and AI-embedded products. This can help to societies. Some argue that in certain circumstances these
reduce unnecessary differences across economies. so called “socio-technical” standards may be even
Addressing such fragmentation is also trade facilitating, implausible, if not impossible (Lin, 2021; Smuha, 2024).25
as it avoids unnecessary compliance costs for companies, However, others consider that such difficulties are not
in particular micro, small and medium-sized enterprises necessarily or always insurmountable and, depending on the
(MSMEs), when engaging in international trade. International specific context and purpose, can be overcome (Ebers, 2024;
standards can be beneficial in other ways. For instance, Kerry, 2024; Meltzer 2023). They argue, for instance, that
they can facilitate the free flow of digital solutions, “foundational” international standards (i.e., those addressing
ensure interoperability, foster innovation by codifying and topics such as terminology, definitions and concepts) may
disseminating best practices in technology (see also Section be less challenging to discuss and adopt than those
4(b)(v) on technology transfer in WTO agreements), shorten addressing substantive or “normative” topics. Indeed, some
the regulatory cycle – as each regulator does not have to foundational AI international standards have already been
start its own process again from scratch, but can benefit adopted.26 Some also note that it may also be possible for AI
from the experience of other regulators – and help small standards to address substantive socio-technical issues (such
companies improve their regulatory compliance. as certain ethical values that an AI system needs to respect),

68
 CHAPTER 4: WHAT ROLE FOR THE WTO?

but only to an extent, that is, not by prescribing in detail cooperation and coordination. Such “soft law” instruments
specifically what ethical AI specifications should be in all can help ensure international standards are better and
cases, but instead by reflecting only general principles that are more appropriately prepared so that they can be a basis for
widely shared across nations (e.g., those reflected in certain designing regulations that can fully attain their policy goals,
international conventions and declarations, such as the UN while at the same time not causing unnecessary obstacles
Universal Declaration of Human Rights).27 to trade. In addition, such decisions and recommendations
support deeper cooperation. The Six Principles are widely
In addition to international standards, some WTO followed by standard-setting bodies seeking international
agreements, such as the TBT Agreement, also promote relevance, and are also recognized in various international
other regulatory coherence tools, such as “mutual and regional fora, as well as in many regional trade
recognition agreements” (MRAs) and “equivalence”. agreements (RTAs) (McDaniels et al., 2018).31
These tools can be useful in facilitating international
trade even when standards, regulations and certification Another example that may be particularly relevant
procedures between trading partners are different or not fully for AI regulation concerns committee guidance
harmonized. Mutual recognition agreements can streamline on conformity assessment (certification). As noted
conformity assessment procedures, allowing economies to above, AI trustworthiness depends on its ability to meet
acknowledge each other’s testing and certification results, stakeholders’ expectations in a “verifiable way”, for example
thereby reducing redundancy, cutting marketing costs and via certification against technical specifications in a regulation
accelerating product dissemination. These agreements can or standard. Conformity assessment procedures are,
help enhance competition and regulatory efficiency, therefore, likely to be key elements in AI regulatory
particularly by opening new markets to foreign access. frameworks.32 In this respect, the TBT Committee’s 2024
Such gains can be significant – a recent study (Cernat CAP Guidelines (WTO, 2024b) stress the need to ensure
2023) indicates that “the existence of an MRA tends to that conformity assessment procedures are “adaptative,
increase the value of exports by 15-40% and the probability responsive, and remain relevant”, which will be instrumental
of firms to export new products to new markets by up to in ensuring safe and trustworthy international trade in
50%”, and states that recent surveys indicate increasing ever changing AI-enabled products. Mutual recognition
interest in economies in tools such as mutual recognition agreements, which as discussed above can help to avoid
agreements “in areas where domestic developments across creating unnecessary trade barriers from duplicative testing
the globe lead to new regulatory requirements”, including and other certification procedures, have also increasingly
in “digital standards, cybersecurity, 5G, interoperability of been the focus of TBT Committee debates and guidance,
electronic invoices and other topics related to the digital including in the CAP Guidelines. The CAP Guidelines build
transformation”. The TBT Agreement, for instance, encourages on the guidance that the TBT Committee has developed
members to rely on equivalence and mutual recognition over the years on “a range of approaches that governments
agreements (Articles 2.7 and 6). Mutual recognition might choose to apply across different sectors to ease the
agreements have been described as important instruments burdens associated with duplicative testing and certification”,
to ensure that unnecessary duplication of certification mutual recognition agreements and equivalence being
procedures does not become itself a barrier to trade on among such approaches.33 In addition, under the Tenth
AI-related products (Meltzer, 2023).28 Triennial Review on the operation and implementation of
the TBT Agreement, a proposal was made for members to
discuss and exchange experiences on the importance and
(iv) Providing voluntary benefits of mutual recognition agreements, including on how
they “may contribute to addressing future global challenges”.34
committee guidance

The WTO also promotes regulatory coherence not

(v) A global forum for
only through the rules of agreements but also through negotiating new rules
“soft law”. An important example is the TBT Committee’s
guidance with respect to international standards. Bearing in
mind the fact that the manner in which international The WTO also promotes global alignment through the
standards are set can have a decisive impact on the extent negotiation of new binding rules on trade. New trade
to which those standards are actually used as a basis for rules are negotiated and agreed to by all WTO members and
convergence, in 2000, the TBT Committee agreed on a set approved domestically. The goal is to ensure that the rules-
of Principles for the Development of International Standards, based international trade system is kept fit-for-purpose, and
Guides and Recommendations (the “Six Principles”).29 The that it provides a level playing field for all, thus contributing to
Six Principles provide guidance in the areas of transparency, economic growth and development.
openness, impartiality and consensus, effectiveness and
relevance, coherence, and development dimension.30 Various issues negotiated under the so called “Joint
Principle 5, on coherence, for example, stresses the Statement Initiative on E-commerce” matter for AI.
importance of avoiding duplication and overlap between The Joint Statement Initiative was launched in January 2017
the work of international standardizing bodies and calls for to respond to the changing nature of trade and create a

69
 CHAPTER 4: WHAT ROLE FOR THE WTO?

modern set of rules to facilitate digital trade and address education, financial and health services. Services that use
challenges within the digital economy. Topics discussed over or rely on AI are often, at least in part, supplied through
the years have included several issues of key importance electronic means. As a result of technological advancements,
for AI, including personal data protection, open government a wide range of services can more easily than previously be
data, access to and use of the internet, cybersecurity, traded across borders as digitized information flows, and
telecommunications, consumer protection, customs duties AI has further increased the tradability of services under
on electronic transmissions, data flows, data localization and mode 1 of the GATS, which refers to the cross-border supply
source code. The negotiations also cover the important issue of services.37 The use of AI by services suppliers may expand
of capacity-building and technical assistance for developing supply capacity and reduce costs. Trade in services also
economies. As of June 2024, 91 WTO members, including stimulates the development and uptake of AI, as access to
many developing economies and several least-developed international markets is a key channel to expand AI-enabled
countries (LDCs), were involved in these negotiations.35 services, monetize the technology and drive investment.

Rules of the GATS carry relevance for AI in these three

key ways. The GATS applies to all services sectors with the
(b) F
 acilitating trade in exception of governmental services (referred to as services
supplied in the exercise of governmental authority) and most
AI-related goods and of the air transport sector. Measures affecting services
services for AI growth supplied through different technological means – e.g.,
electronically or with the assistance of AI – are all covered
and development by the GATS.38 While certain obligations of the GATS apply
to all services within its scope (e.g., the obligation to publish
measures of general application), some of the principal
The WTO is the cornerstone of global efforts obligations do not apply to all services covered. For example,
to facilitate trade in services and in goods that market access (Article XVI) and national treatment (Article
enable or are enabled by AI. The expansion and XVII), the two obligations that aim to guarantee a level of
development of AI, and its increasing use by firms and openness to international competition, only apply to those
individuals around the world require a facilitating trade and services sectors that are listed in the schedule of specific
investment environment. WTO agreements encourage commitments of each WTO member, and in accordance with
policies contributing to a sound environment for investment limitations listed for particular modes of supply.
and cross-border trade in AI-related products and
technologies. Various aspects of the WTO rulebook can Most WTO members have made specific commitments
contribute to promoting the development of and access to AI. on market access and national treatment for
computer services. Out of the WTO’s (counting the
European Union as 1) 141 schedules of commitments, 84 (or
(i) Obligations and 60 per cent) contain commitments on computer services, but
specific commitments only 53 contain specific commitments covering the totality
of the sector as defined in the GATS classification system.39
on trade in services In addition, the level of treatment bound for each mode of
supply varies. Of specific commitments in the different sub-
sectors of computer services, 67 per cent were unrestricted
The obligations of the General Agreement on Trade in (i.e., without sector-specific limitations) for cross-border
Services (GATS) play an important role in shaping a supply (GATS mode 1), in comparison with 74 per cent
policy environment that facilitates the development for consumption abroad (GATS mode 2), and 64 per cent
and uptake of AI. AI is relevant for trade in services – for commercial presence (GATS mode 3).40 For their part,
including trade in services for AI – in three key ways. First, commitments on GATS mode 4 are typically limited to certain
while AI has many different applications, the development categories of natural persons, notwithstanding the sector.
and implementation of AI is, at its core, a computer service.
In the sectoral classification system used under the GATS, Subsectors of telecommunication services have a
computer services comprise a wide range of services relating higher number of commitments. A total of 100 schedules
to the design and development of computer systems and contain commitments in the sector (including 43 that include
software.36 Computer services under the GATS also include commitments across all subsectors),41 but those tend to be
data processing and database services, which are key subject to a higher number of limitations for both modes 1
functions associated with AI, given its high level of reliance on and 3. For example, 67 per cent of commitments on data
access to, and treatment of, data. Second, telecommunications transmission are subject to limitations or are “unbound” with
services play a fundamental role in enabling and promoting respect to mode 1, and the proportion of commitments with
AI. AI relies on efficient communications infrastructures limitations reaches 79 per cent for mode 3.
to provide the levels of connectivity it requires to function,
including by facilitating the transmission of data within and However, commitments in other sectors remain
across borders. Third, AI is used as an input in the supply of limited, making for a less predictable and transparent
an increasingly wide range of services, including translation, trade environment in these sectors. As noted above, AI

70
 CHAPTER 4: WHAT ROLE FOR THE WTO?

is used as an input in the supply of a wide range of services, shows the proportion of schedules with specific commitments
where commitments have relevance, including under mode 1. under mode 1 for a sample of subsectors. The absence
Overall, commitments under the GATS are limited, as most of specific commitments means that no guarantees of access
sectors attract fewer commitments than the computer and are provided, and this makes for a less predictable
telecommunications sectors. Indeed, a majority of WTO and transparent trade environment for the relevant sectors,
members have not scheduled commitments in most of the as new trade-restrictive measures may be imposed at
sectors covered by the GATS. On average, WTO members’ any time. The limited multilateral commitments in different
schedules have specific commitments in roughly a third of sectors also represent a lost opportunity to encourage lower
all services subsectors. In addition, even when commitments levels of services trade restrictiveness. Indeed, some of the
are undertaken, many services subsectors have been left services sectors of greatest relevance for AI remain subject
unbound (i.e., free to limit both market access and national to significant trade restrictions, applied by different
treatment) for mode 1. This is illustrated in Figure 4.1, which governments around the world.

Figure 4.1: Proportion of GATS schedules with specific commitments in modes 1 (cross-border supply)
and 3 (commercial presence) in selected sectors

% 10 20 30 40 50 60 70 80

Legal

Accounting/auditing

Software implementation

Data processing

R&D - natural sciences

Advertising services

Voice telephony

Packet-switched data transmission

Online information and data retrieval

Audiovisual services

Retailing services

Educational services

Non-life insurance

Acceptance of deposits

Health-related and social services

Travel agencies and tour operators

Freight transport agency

Mode 1 Mode 3

Source: WTO using I-TIP Services ([Link]

71
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Figure 4.2: Services trade restrictiveness in selected sectors

0 10 20 30 40 50 60

Transport

Tourism

Telecommunications

Professional

Health

Finance

Distribution

Construction

Computers

Source: WTO using I-TIP Services ([Link]

Note: This chart depicts the average level of restrictiveness in the applied regimes of 133 economies in nine broad sectors. The index
quantifies applied services trade policies on a scale from 0 (fully open) to 100 (most trade-restrictive).

Aside from the level of treatment guaranteed by In addition, 58 economies have Services Trade Restrictiveness
commitments, barriers to services trade actually Index (STRI) scores of 50 or above for either mobile or
applied by governments remain high in overall terms. fixed-line telecommunications. Restrictions are also important
However, these barriers display significant variations across in a number of services sectors that use AI, including
sectors, modes of supply, regions and levels of development financial services, which thereby limits capacity to supply
(see Figure 4.2). Sectors such as professional and transport AI-intensive services and impacts growth opportunities.
services, for example, tend to be more restricted than
telecommunications, computer or distribution services. Restrictions in computer and telecommunications
services are highest for mode 1 (cross-border supply)
Services sectors particularly crucial to AI, such as and significant for modes 3 (commercial presence)
computer services and telecommunications services, and 4 (movement of natural persons) (see Figure 4.3).
still face significant trade restrictions in a large Restrictions in mode 1 may affect the cross-border supply
number of economies. With respect to computer of consultation services relating to computer systems and
services, 24 economies (out of a sample of 133) have software, which are important for the development of AI and
services trade restrictiveness scores of 50 or above on its implementation and use in companies. Mode 1 restrictions
a scale from 0 (fully open) to 100 (most trade-restrictive). on computer and telecommunications services can limit

Figure 4.3: Services trade restrictiveness by mode of supply

a) Telecommunication services b) Computer services

80 80

70 70

60 60
STRI: Computer
STRI: Telecom

50 50

40 40

30 30

20 20

10 10

0 0
M1 M3 M4 M1 M3 M4
Mode of supply Mode of supply Mode of supply

Source: World Bank and World Trade Organization (2023).

Note: Figure 4.3(a) and (b) depict average Services Trade Restrictiveness Index (STRI) scores by mode of supply for 133 economies.
The STRI quantifies applied services trade policies on a scale from 0 (fully open) to 100 (most trade-restrictive). M1: GATS mode 1 –
cross-border supply; M3: GATS mode 3 – commercial presence; M4: GATS mode 4 – movement of natural persons.

72
 CHAPTER 4: WHAT ROLE FOR THE WTO?

the transmission of data and cross-border data processing efficient infrastructure for services contributing to, or using,
and storage activities. Mode 3 restrictions have particular AI. For example, the Annex provides for access to public
significance, as they include measures that affect the capacity basic telecommunications services on reasonable and
of foreign suppliers to establish a commercial presence non-discriminatory terms and conditions for the supply of
abroad, and to supply services through such commercial services in all committed sectors. It also mandates that
presence. When applied to computer services, restrictions suppliers from other members should be able to use public
to mode 3 impede foreign companies from investing and basic telecommunications services to enable the flow of
being active in the local market for AI and related services. information within and across borders.
Restrictions to mode 3 in telecommunications services
limit investment in the digital infrastructure that is critical Newly agreed disciplines on services domestic
to enable the movement of data and the electronic supply regulations and investment facilitation, which aim
of a wide range of services, including those relying on AI. to improve the business environment, can also help
As for limitations to mode 4, these encompass measures that to facilitate the development and use of AI. The
affect the capacity of experts who work on the development disciplines on services domestic regulation, which entered
of AI systems and software to temporarily go abroad to supply into force in February 2024, facilitate authorization
these computer services. procedures that businesses engaged in AI-related or AI-
enabled services may have to comply with before supplying
Overall, services trade restrictions raise trade costs their services in various jurisdictions (WTO, 2024). A total
and limit trade and investment. They carry negative of 72 governments, representing 92.5 per cent of global
economy-wide consequences and worsen the performance services trade, have committed to implementing these new
of the specific sectors targeted (World Bank and WTO, 2023). disciplines, which will be applied on a “most-favoured-
In the case of telecommunications services, for example, nation” basis, meaning they will benefit all WTO members.
trade restrictions have been associated with lower WTO members that have adopted the disciplines on
penetration, higher prices and lower-quality services services domestic regulation have embraced good regulatory
(Borchert et al., 2017; ITU and UNESCO, 2013; Nordås practices on stakeholder involvement: these practices
and Rouzet, 2017). Meanwhile, trade restrictions in relation foresee the advance publication of draft laws and
to digitally supplied services limit an economy’s capacity regulations relating to licensing, qualifications and technical
to take advantage of trade opportunities created by AI standards. They also foresee that interested persons
and technological developments, and can also reduce are given reasonable opportunity to comment on such
companies’ incentives to invest in digital technologies and in draft regulations, and the consideration of such comments
information and communications technology (ICT). by the regulators. In addition, the recently completed
Agreement on Investment Facilitation for Development,
In addition to the market access and national treatment concluded by close to 130 members, aims to improve
obligations, the GATS contains other obligations the investment and business climate and make it easier
which generally aim to facilitate services trade. for investors to conduct their day-to-day business and
These obligations can affect the trade policy environment expand their operations. Although this is a plurilateral
for AI and the propensity of AI to increase services trade. agreement, its benefits would extend to all members.
In addition to the most-favoured-nation obligation (Article With incorporation into the WTO architecture, this agreement
II) and transparency requirements (Article III), Article VI will also help to attract more and higher-quality investment
contains obligations on domestic regulation that require, in digital connectivity infrastructure. Such infrastructure
among other things, the reasonable, objective and impartial forms the backbone for deploying digital technologies,
administration of measures in sectors in which specific including AI.
commitments are undertaken. Several WTO members
have also included additional commitments on domestic
regulation in their schedules by means of a reference paper (ii) C
 ustoms duties on
containing disciplines that seek to mitigate the unintended
trade-restrictive effects of measures relating to licensing
ICT equipment and
requirements and procedures, qualification requirements electronic transmissions
and procedures, and technical standards.

The telecommunications sector – a key enabler of AI, Tariffs, especially on ICT equipment, can limit access
data flows and digitally delivered services using AI – to and increase the cost of hardware essential to
is also the focus of two additional sets of develop and power AI applications. They can thereby
competition-related rules under the GATS. These rules constitute an obstacle for the deployment and adoption of
are the Annex on Telecommunications, which applies to all AI technologies. Acknowledging the growing importance
WTO members, and the Reference Paper on Regulatory of ICT products to promote competitiveness in the digital
Principles on Basic Telecommunications, which has been economy, a subset of WTO members negotiated an
incorporated into the Schedules of Commitments of 103 agreement – the Information Technology Agreement (ITA) – to
WTO members. By promoting competitive conditions eliminate tariffs on such products. Beyond tariffs, WTO rules
and good regulatory practices in the sector, the two also provide a vehicle to determine the value for AI-enabled
instruments help to foster the extension of affordable and goods (see Box 4.1).

73
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Box 4.1:
AI and customs valuation

The incorporation of advanced government officials and traders design work, if supplied by the
digital technologies, including AI, alike in valuing goods, and growth buyer and not undertaken in the
into products creates challenges in AI-enabled products could importing country (Article 8.1(b)).
for governments seeking to potentially add to the uncertainty These provisions could be relevant
determine the value of those relating to national valuation when determining the value of
products for tariff and other practices and the extent of revenue AI-enabled products. Moreover,
purposes. For decades now, collection at the border. WTO members may elect whether
customs agencies have grappled to include the value of software in
with how to determine value Customs valuation rules can be a certain “carrier media” (i.e., physical
for imported goods that bundle vehicle to capture the value of the devices bearing the software),
hardware and software elements. AI-enabled features of imported although this discretion is limited
Customs valuation is primarily goods. The WTO Customs to devices that exclude integrated
concerned with the transaction Valuation Agreement allows circuits or semiconductors and
value of physical goods, from WTO members, under specified therefore may not extend to certain
which accompanying services circumstances, to value certain advanced digital technologies
or elements may be excluded. intangibles embedded in imported that feature AI.42 The challenge of
While there is some scope for products. The transaction value mapping existing rules onto new
determining the value of certain of goods can be augmented with market developments could be
intangibles associated with such elements in certain instances, particularly acute when dealing
imported products, determining for example, where there are IP with the fast-changing
whether declared value accounts, royalties or licence fees (e.g., developments in AI-enabled
or should account, for these patents, copyrights and trademarks) products (see Chapter 4(f)).
intangibles can entail complex related to the goods and tied to
considerations and can lead to their sale (Article 8.1(c) of the At the same time, the use of AI,
exchanges between customs Customs Valuation Agreement), including predictive AI models,
agencies and importers to or where the production of imported has significant potential to change
verify certain elements of the goods has been dependent on the work of customs officials
transaction. This has been such items such as the cost of when valuing imported products
a persistent challenge for engineering, development and (see Chapter 2(b)).

The ITA aims to increase worldwide access to high-tech Tariff rates on ICT products by non-ITA participants
goods, such as semiconductors, which are essential are highest for low-income and lower middle-income
to AI, by eliminating tariffs on ICT products covered economies. This limits the capacity of these economies to
by this Agreement. Participation in the original ITA has leverage AI for development. Tariffs rates vary significantly
increased from 43 WTO members in 1996 to 84 today, across levels of development. While they average 6 per cent
representing about 97 per cent of world trade in IT products. in high-income and upper middle-income economies, they
In 2015, over 50 WTO members, including China and the reach almost 8 per cent in lower middle-income economies
United States, concluded the expansion of the original and 9 per cent in low-income economies (see Figure 4.4).
agreement (ITA II), which covers an additional 201 products.
ITA commitments to provide duty-free access to ICT products Beyond the ITA, the WTO moratorium on customs
are applied on a most-favoured-nation (MFN) basis, that is, duties on electronic transmissions can contribute to
to all WTO members, including non-ITA participants. The promoting access to AI. The moratorium, which ensures
value of products covered by the ITA II reached US$ 2.1 that no tariffs are imposed on electronic transmissions,
trillion in 2021. The elimination of tariffs on products such as and has been periodically renewed since 1998, ensures that
semiconductors promotes access to hardware that is essential additional costs are not imposed on electronic transmissions
to power AI systems. As noted in Chapter 2, demand for AI in the form of customs duties. The last extension of the
hardware components, such as CPUs, GPUs and specialized moratorium was agreed in March 2024 at the WTO’s 13th
AI chips, has been rising sharply. The ITA II also contains a Ministerial Conference (MC13). WTO members agreed to
commitment to keep the list of covered products under review renew the moratorium until the 14th Session of the Ministerial
to determine whether further expansion may be needed to Conference or 31 March 2026, whichever is earlier. The
reflect future technological developments. Ministerial Decision notes that “the moratorium and the

74
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Technical regulations, standards and conformity

Figure 4.4: Tariffs on ICT products by non-ITA assessment procedures are subject to the WTO TBT
participants (2023)* Agreement. This agreement supports better regulatory
systems, which are essential for ensuring AI trustworthiness,
Low-income
10
and, through this, the deployment of AI. The TBT Agreement
provides a framework of disciplines related both to
9 procedural (transparency) and to substantive (product
8 specifications and certification) dimensions of regulatory
processes, which are aimed at eliminating unnecessary or
MFN simple average (%)

7
discriminatory technical barriers to trade, while safeguarding
6 the right to regulate to address legitimate policy
objectives (see also chapters 4(a), 4(c), 4(d), and 4(e)).
5
TBT-compliant regulatory measures are important for the
4 conduct of international trade, including trade in AI systems
3 and AI-enabled products, because they can increase
consumers’, importers’, and other stakeholders’ trust in the
2 safety and quality of the traded products. This can help to
1 ensure that trade flows smoothly, while respecting the right
of governments to regulate for legitimate policy reasons.
0
High- Upper Lower Low- This trust does not however arise spontaneously. Instead,
income middle- middle- income
group income income group “behind the scenes”, trust is supported by an “invisible chain”
group group of institutions working together to deliver what is referred
to as the National Quality Infrastructure (NQI), a normative
*2022 for Saint Kitts and Nevis, Democratic Republic of
the Congo, and Haiti; 2019 for Yemen; and 2016 for Djibouti and institutional framework composed of a combination
(latest year available). of regulations, standards and certifications, as well as
Source: WTO Analytical Database. agencies, laboratories and other facilities that are responsible
Note: Product codes S04, T03, T04 and T05 of the for applying these measures (WTO, 2021; 2024b). As trust
multilateral trade negotiations product categories. increasingly underpins AI deployment and use, the role of
the NQI will also increase in this area.46

Work Programme will expire on that date”.43 Members have (iv) Agreement on
expressed differing views concerning the renewal of this
temporary moratorium.44 The non-imposition of customs
Trade-Related
duties on electronic transmissions is part of the Joint Aspects of Intellectual
Statement Initiative on E-commerce text (see above).
Property Rights
(TRIPS Agreement)
(iii) T
 echnical Barriers to
Trade (TBT) Agreement The WTO TRIPS Agreement, the most comprehensive
multilateral agreement on IP, directly impacts the
development, deployment and commercialization
Governments, civil society and economic operators of AI technologies. Established in 1994, the TRIPS
broadly agree on the pivotal role of mandatory technical Agreement sets down minimum standards of protection and
regulations, voluntary standards and conformity enforcement for IP rights across WTO members. It outlines
assessment procedures in ensuring that AI systems the obligations of members to protect IP, including with
are trustworthy.45 This is essential to promote the regard to copyrights, patents, trademarks, industrial designs
deployment of AI. Technical regulations and standards and trade secrets, all of which are relevant to AI technologies
are used to set out specifications and requirements on and AI-generated creations and innovation.
the production, importation and sale of products. As such,
when adopted and applied appropriately, they can provide The TRIPS Agreement envisages a balanced IP
an essential regulatory framework for the development and system that not only incentivizes innovation but
use of trustworthy AI systems, and can ensure that risks also promotes access to and dissemination of
associated with AI are addressed and that its benefits are technology. By means of this system, the enforcement
harnessed. To ensure that the policy goals pursued by such and protection of IP rights contribute positively to
measures are fully attained in practice, economies also technological innovation and to the mutual benefit of both
need to subject AI systems, including AI-enabled products, producers and users of technological knowledge, thereby
to conformity assessment procedures in order to assess supporting social and economic welfare. This objective
whether relevant requirements for ensuring trustworthiness is fundamental for the development and application of AI
have been fulfilled. in the future.

75
 CHAPTER 4: WHAT ROLE FOR THE WTO?

The minimum requirements for IP protection required important role in promoting the development of AI.
by the TRIPS Agreement can serve to address The TRIPS Agreement as a whole pursues the objective
certain IP challenges arising from the development that the protection and enforcement of IP rights should
and applications of AI, albeit with some limitations contribute to the promotion of technological innovation
and challenges. As set out in Chapter 2(b), IP rights are and to the transfer and dissemination of technology
relevant to the development of AI, including the use of its (Article 7), as a balanced and reliable IP system can
inputs and the protection of its outputs. Disclosure provide the legal infrastructure through which intangible
requirements under international patent rules can result in assets and knowledge can be traded. In addition, Article 8
a positive contribution to transparency in the development of the TRIPS Agreement underscores the principle that
of AI technology. Under the TRIPS Agreement, patent such IP protection is not inconsistent with members
applications require the applicant to disclose the invention pursuing public interest considerations. Article 8 also
in a manner sufficient to enable a person with the relevant acknowledges that members may need to take appropriate
skills to replicate the invention.47 Where jurisdictions provide measures to prevent the abuse of IP rights by right-
patent protection for software or computer-implemented holders or the resort to practices which unreasonably
inventions, this disclosure requirement yields significant restrain trade or adversely affect the international
expert information on patented technologies generally and transfer of technology. The TRIPS Agreement also
can be used to address the “black box” problem that may mandates developed members to provide incentives
arise with AI (see Chapter 2(a)), at least to a certain extent. to their enterprises and institutions for the purpose of
promoting and encouraging technology transfer to LDCs.50
Under Article 10 of the TRIPS Agreement, computer The TBT Agreement, which encourages the use of international
programmes, whether in source or object code, standards as a basis of regulations, expressly recognizes
are protected as literary works under the Berne “the contribution which international standardization can
Convention (1971). This robust protection for software make to the transfer of technology from developed to
under copyright may provide a further incentive for developing countries.”51 Article IV of the GATS encourages
transparency and to publish AI algorithms rather than the increasing participation of developing economies in
keeping them protected as trade secrets. Nevertheless, the world trade through the negotiation of specific commitments
TRIPS Agreement also requires WTO members to protect to build domestic capacity, efficiency and competitiveness,
undisclosed information, including trade secrets, under including through access to technology on a commercial
legislation against unfair competition (Article 39 of the TRIPS basis. And a Working Group on Trade and Transfer of
Agreement). Ultimately, the attribution of IP rights in principle Technology was established at the Doha Ministerial
does not determine whether their exercise is restrictive or Conference in 2001 with the aim of examining the relationship
permissive, and open-source solutions may be encouraged between trade and transfer of technology from developed
by regulation if deemed desirable by policymakers. to developing economies and ways to increase this flow
of technologies.52
IP rights also provide the legal framework to determine
the rights of creators whose works and/or databases Several technology transfer programmes relevant for
are used as input to train AI. Regarding exceptions to IP AI have been reported in recent years. Since 2019,
rights, including “fair use”, the TRIPS Agreement introduces in the context of the TRIPS Council, a few developed
a three-step test48 that establishes the criteria for members economies, including Canada, the European Union,
to follow when they establish exceptions and limitations to Switzerland and the United States, have reported that they
IP protection, such as text and data mining for training and adopted several relevant AI technology transfer programmes
developing AI models. in order to fulfil their commitments to incentivize local
enterprises to promote and facilitate technology transfer to
Finally, with regard to the issues of AI output, the LDCs, with the aim of helping these LDCs establish a sound
TRIPS Agreement establishes minimum standards. and viable technological base.53
While it is based on the traditional, human-centric approach
to IP, it does not preclude members from addressing However, the extent to which technology transfer
issues arising from new technologies in their domestic provisions have been used is a subject of debate.
legislation. In addition, the flexibilities included in the Research indicates that the implementation of Article
TRIPS Agreement allow WTO members to implement their 66.2 of the TRIPS Agreement has been uneven and
obligations in a manner consistent with their own legal that the reporting by developed economies on their
system and developmental needs. The TRIPS Agreement obligations has often been inadequate or lacking in detail
can, therefore, be used to address AI-related IP issues in (Moon, 2008). Developed economies argue that, in most
tailored approaches.49 cases, IP is in the hands of the private sector, which
makes it difficult to transfer technology. Developing
(v) Technology transfer in WTO members, on their side, question the extent to
which these provisions have effectively encouraged
WTO agreements technology transfer and benefited developing economies.54
It has also been noted that the best-endeavour formulation
of these provisions, which do not set any clear
Various WTO agreements include provisions to mechanisms or tools for technology transfer, hinders the
promote technology transfer, which can play an implementation of the disciplines (Mishra, 2024).55

76
 CHAPTER 4: WHAT ROLE FOR THE WTO?

(vi) Agreement policy objectives – related, for example, to health, environment

on Government or safety – without creating unnecessary or discriminatory

technical barriers to trade. The TBT Agreement therefore
Procurement (GPA) provides ample policy space to regulate AI, while preferring
interventions that are non-discriminatory and are the least
trade-restrictive possible to fully achieve the stated legitimate
The rules of the WTO GPA 2012 promote access policy objectives. Attaining this regulatory balance can
to internationally available new AI technologies. help to ensure that trade flows smoothly, while respecting
The GPA 2012 aims to open up, to the extent agreed governments’ right to regulate for legitimate policy reasons.
by parties to the Agreement, government procurement This can also be important in ensuring that discriminatory
markets to suppliers from other GPA parties, and to make or unnecessarily burdensome standards and regulations
government procurement more transparent and predictable. do not hamper interoperability of AI systems and products
It provides legal guarantees of non-discrimination for the (Lim, 2021).57
goods, services and suppliers of GPA parties with regard
to government procurement covered by the Agreement, The principle of a periodic review of standards,
including of AI tools, as the case may be. The Agreement regulations and certification procedures enshrined
does not contain any direct references to AI. However, it in the TBT Agreement is particularly suitable for
does require that GPA parties, where appropriate, set fast-evolving technologies such as AI. The TBT
out technical specifications in terms of performance, and Agreement requires that regulations shall no longer be
functional requirements and base technical specifications maintained, or that they shall be updated, in light of changes
on international standards, where such standards exist, in the circumstances that gave rise to their adoption.58
or otherwise on domestic technical regulations or recognized Members are encouraged to evaluate their regulations
domestic standards. Moreover, AI technologies can be used periodically so as to ensure that they are fit-for-purpose
to implement the GPA, such as by identifying red flags that as technological and other circumstances evolve over
might point to corrupt practices or conflicts of interest or time. For instance, new scientific or technical evidence on
collusion, and by collecting the relevant statistical data. the risks and challenges of AI, or other circumstances that
Reflecting the growing importance of AI tools procurement, led to the adoption of an AI standard or regulation may
some GPA parties, including the European Union, have become available after their adoption. Depending on the
published standard contractual clauses to be used by its nature and extent of such new developments, this may
procuring entities when purchasing AI tools. require updating and recalibrating the measure accordingly.
Regulations may also need to be revised to take account
of any a new or revised relevant international standards.59
As already noted, the TBT Agreement requires standards
(c) M
 inimizing negative and regulations to be based on relevant international
standards. The importance of periodically evaluating and
international spillovers revising international standards in light of relevant changes,
such as new scientific and technological developments,
to prevent them from becoming obsolete, is expressly
The WTO rulebook includes various principles, mentioned in the TBT Committee’s Six Principles (Principle
provisions and guidelines that can support the 4). The principle of periodic review is also underscored in
deployment of AI, as well as trade in AI systems and the TBT 2024 conformity assessment procedures (CAP)
AI-enabled products, by minimizing negative guidelines (WTO, 2024b). Building on TBT Agreement
international spillovers. For example, non-discrimination, provisions on this issue,60 the CAP guidelines expressly note
a key principle of the WTO, is meant to prevent that “the choice of the conformity assessment procedures
discriminatory treatment of foreigners and trading partners.56 should not be seen as permanent. It should benefit from
Another example is the Agreement on Trade-Related regular review as the elements that influenced the original
Investment Measures, which recognizes that certain choice of conformity assessment procedure may change
investment measures can restrict and distort trade and states over time.”
that WTO members may not apply any investment measure
that discriminates against foreign products or that leads to Such approaches are important from both a policy
quantitative restrictions. and trade perspective. Regularly updating standards,
regulations and certification procedures helps to maintain
The TBT Agreement provides that regulatory their effectiveness in addressing their intended policy
intervention shall not be discriminatory, nor more goals (such as health or safety) even when the features,
trade-restrictive than necessary to achieve the characteristics and risks of what they regulate, including
intended policy objectives. When it comes to AI, evolve over time. But this can also be beneficial from
technical regulations, voluntary standards and certification a trade perspective when changes in the circumstances
procedures, which play a critical role in ensuring AI systems giving rise to the adoption of a regulation open new
are trustworthy, the TBT Agreement aims to ensure that alternatives for redesigning it so that it can still fully attain
regulatory measures are prepared, adopted and applied in its policy objectives, but in a less burdensome, trade
such a way that they can both fully attain their legitimate restrictive way (Lim, 2021).61

77
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Work has also been carried out on how “undue trade effects. Subsidized products can also be the subject of
distorting effects” of non-tariff measures (NTMs) countervailing measures applied by an importing member,
in ICT products could be reduced or eliminated if the subsidized imported goods are found to cause
to prevent such measures potentially offsetting injury to the importing member’s domestic industry producing
ICT tariff market access gains. Such NTMs include the same or similar goods.
technical regulations, certification procedures and labelling
requirements. In November 2000, the ITA Committee Where the product incorporating AI is a good, the
approved a work programme on this topic that SCM Agreement and the actions and remedies
resulted in the adoption, in February 2005, of the described above apply only to subsidies that are
Guidelines for Electromagnetic Compatibility (EMC) specific. A subsidy may be considered “specific” if access
and Electromagnetic Interference (EMI) Conformity to it is explicitly limited to a particular enterprise, industry,
Assessment Procedures (“EMC/EMI CAP Guidelines”).62 group of enterprises, group of industries, or a specific region.
Following adoption of these guidelines, the WTO This fact could be pertinent for broad AI initiatives that,
Secretariat was asked to compile information on the at least to some extent, involve goods. In particular, it is
different types of conformity assessment on EMC/EMI. important to consider whether a government financial
This information has since then been updated regularly support programme for AI is available to a wide range
(WTO, 2017).63 of economic activities or is more narrowly targeted at
particular sectors or enterprises. For instance, it could be
WTO disciplines on subsidies in the Agreement challenging to identify specificity in a government subsidy
on Subsidies and Countervailing Measures (SCM intended for general AI development and which could be
Agreement) can also play a crucial role in navigating utilized in diverse sectors, such as healthcare diagnostics
the dual aspects of AI development: promoting and autonomous driving systems.64 Such a subsidy might
technological innovation while preventing negative spillovers appear to support broad technological advancement
in international trade resulting from government financial (thus, potentially non-specific), while in practice it
support. As outlined in Chapter 3(b), an increasing disproportionately benefits certain industries or companies
number of governments is implementing AI strategies engaged in specific commercial activities involving goods
with significant financial components and putting in place that incorporate AI (thus, potentially specific). The specificity
strategies to promote access to data. The relevance and analysis also may be complicated by the rapid evolution
applicability of the WTO subsidies disciplines to prevent and dual-use nature of AI technologies. Such ambiguities
negative spillovers relating to government financial make it difficult to generalize; any assessment of specificity
support for AI or to the provision of data by government necessarily depends on the particular facts of a given
as an input depend on numerous elements. First is the situation. The ambiguities regarding specificity can lead to
nature of the traded product and whether it is considered differing views among trading partners as to the actionability
a good or a service. The SCM Agreement does not apply of certain subsidies, where some trading partners may be
to services or IP as such, but instead exclusively applies concerned that subsidies provided by others are unfairly
to goods. Consequently, it is essential to distinguish distorting international competition.
hardware components and AI-enabled products that are
classified as goods (to which the SCM Agreement would Subsidies directed toward the production of AI
apply) from AI software itself. To the extent that the integrated hardware or AI-enabled goods may
AI component in any given good – for example, the AI in present less ambiguity regarding their specificity.
an autonomous vehicle or in advanced robotics – benefits For example, a subsidy might be provided for the production
from subsidies coveredby the SCM Agreement (the SCM of advanced sensors that are explicitly used in both
Agreement defines a subsidy as a financial contribution by commercial drones and military surveillance equipment.
a government or public body or any form of income or price The targeted nature of such a subsidy to the production
support that “confers a benefit” on the recipient), further of a certain limited set of goods could make it easier to
analysis may be required to determine whether these identify the subsidy as specific under the SCM Agreement.
subsidies could be attributed to those goods, and thereby A further aspect of specificity, as mentioned above, is
could become the subject of counteractions under the that the SCM Agreement deems as specific the two
SCM Agreement. categories of prohibited subsidies: those contingent on
export performance, and those contingent on the use of
Subsidies may be challenged in WTO dispute domestic goods over imported ones, commonly referred
settlement under the SCM Agreement. If the subsidy to as import substitution subsidies. It should be noted
in question is a prohibited subsidy (such as an export here that while import substitution subsidies are prohibited,
subsidy, or a subsidy for the use of domestic goods rather subsidies supporting exclusively domestic production are
than imported goods), or if it causes serious prejudice or not prohibited. Nevertheless, to the extent that a subsidy
other specified adverse effects to another member’s trade of the latter type is specific, it could be the subject of
interests, a multilateral remedy to offset the harm can counter actions provided for in the SCM Agreement, i.e.,
be authorized through the WTO. In cases of prohibited through WTO dispute settlement or the application of
subsidies, the remedy requires the withdrawal of the countervailing measures. These points highlight the need
subsidy. For actionable subsidies, the remedy involves either for awareness of the rules of the SCM Agreement when
the withdrawal of the subsidy or the removal of its adverse designing subsidy programmes for AI.

78
 CHAPTER 4: WHAT ROLE FOR THE WTO?

technologies, including AI. For instance, cybersecurity,

(d) Helping to address an increasingly important consideration in AI regulations
and policies (see Box 4.3), is a common theme of various
and prevent STCs. More directly on AI, from March 2022 to June 2023,
trade tensions the TBT Committee discussed a concern raised with respect
to the EU AI Act (AIA), the first broad regulatory measures
and frictions on AI systems (be they standalone or embedded into
physical products, e.g., a toy). Among other matters, this
concern entailed issues related to the scope and meaning
The practice of raising specific trade concerns (STCs) of the definition of “AI system” and the possibility that
and the requirement to notify technical regulations regulatory authorities could be granted access, as part of
at a draft stage can help to defuse potential trade the certification process, to source code of AI systems.65
tensions. Members commonly use WTO bodies to raise Beyond AI, STCs have also been raised in relation to other
specific trade concerns with respect to laws, regulations, technologies, such as IoT and robotics, which are often used
or practices by their trading partners which may affect in tandem with, or may embed, AI (see Box 4.3).
their trade (see Box 4.2). Since 1995, members have
devoted an increasing amount of time and attention to The WTO also serves as a global forum to settle trade-
discussing STCs. These discussions can help to ease trade related disputes. One of the key functions of the WTO is
tensions by providing members with further information to ensure the integrity and respect of trade rules by providing
and clarification on the rationale behind other members’ a formal system for handling the settlement of trade disputes
regulations, enabling them to work towards mutually among WTO members. A member may bring a dispute to
satisfactory solutions and helping to build trust (see the the WTO’s Dispute Settlement System to seek the redress
opinion piece by Dan Trefler). As noted in Chapter 4(a)(ii), of a violation of obligations or other nullification or impairment
the TBT Agreement also requires members to notify draft of benefits under the WTO agreements or an impediment to
regulatory measures. This requirement can help to defuse the attainment of any objective of the WTO agreements.66
tensions at an early stage, before a measure is adopted Reports by adjudicators specifically selected for a given
(Lim, 2021; Possada et al., 2022). dispute (called “panels”) are considered for adoption by the
Dispute Settlement Body (DSB), that is, all WTO members.
Members have been using the STC practice in These reports are limited to the specific legal and factual issues
the TBT Committee to discuss regulations and raised in the dispute. Many disputes are settled through
conformity assessment procedures on various digital consultations even before any decision is rendered.67

Box 4.2:
The practice of specific trade concerns

STCs, which drive the detailed, members to reduce potential Karttunen, 2020; Lim, 2021; World
technical deliberations on specific trade tensions effectively, and Trade Organization, 2020b).
measures that have, mostly, not yet in a cooperative, non-litigious
entered into force and are therefore manner. This practice thus creates Evidence suggests this model works.
not yet entrenched in domestic opportunities for regulatory While, since 1995, around 56,000
law, can contribute to an improved cooperation centred on a “peer regulatory measures have been
understanding by members of the to peer learning” process, in notified to the TBT Committee, only
rationale underlying other members’ which critiques are presented, around 830 STCs been raised and
regulations. They can also present suggestions are posited, technical, discussed, with even fewer formal
an opportunity to question the legal and policy arguments are disputes (11) involving TBT measures
appropriateness or effectiveness of made, and regulatory experiences having been adjudicated.68 Even if
trade measures, including in terms are exchanged on specific it is not perfect, and there is room
of their scientific or technical basis regulations addressing real for further improvement (Holzer,
or the evidence for them, use of life issues. This provides a 2019), the practice of raising and
international standards, transparency, collaborative “space for learning discussing TBT STCs is generally
and possible regulatory alternatives. from differences” (OECD/WTO, accepted to be a success
2019), which can ultimately lead (Karttunen, 2020) – one that could
Raising concerns via an open, to more effective regulatory be expanded into other WTO
multilateral platform can help outcomes (Horn et al., 2013; committees (Possada et al., 2022).

79
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Box 4.3:
TBT, AI, the Internet of Things and robotics

In view of the significant benefits notifications concern this broader consensus around the pivotal role
and challenges that the Internet group of digital technologies, i.e., that international standards can play
of Things (IoT) and robotics can measures addressing IoT and in ensuring interoperability.74 Specific
engender, in particular when “smart functionality” (19),69 discussions on IoT and robotics-
enabled by AI systems (Suleyman autonomous vehicles (18),70 related standards and policies are
and Bhaskar, 2023), they have robotics (16)71 and industrial taking place in international bodies
increasingly become the object automation (18).72 In addition, and organizations, such as ASTM
of governmental regulatory under this broader group, five International, the International
interventions and policies. In this STCs have been raised concerning Electrotechnical Commission (IEC),
respect, a growing number of IoT IoT/robotics-related measures.73 the International Organization
and robotics-related measures have for Standardization (ISO), the
been notified to the TBT Committee. Interoperability, which is key for International Telecommunication
connecting infrastructures and Union (ITU), the Organisation
These notifications are part of a systems and deploying IoT and for Economic Co-operation and
broader context, in which WTO robotics (WTO, 2018), is among the Development (OECD) and the
members are increasingly notifying a issues addressed in some of these United Nations Economic
wide range of regulations on digital notifications and STCs. As it is the Commission for Europe (UNECE),
technologies to the TBT Committee case with most digital technologies, most of which are observers to the
(Lim, 2021). To date, at least 71 TBT including AI, there is general TBT Committee.

The importance of enforcing legally binding rules on the digital divide and social inclusion and access to
AI at a global level has been highlighted in information is a reasonably important policy objective” and
international initiatives. For example, ensuring found that the measure at issue was at least “designed” to
compliance and accountability based on norms is one of protect “public morals” within the meaning of the general
the seven institutional functions identified in the UN exception under Article XX(a) of the General Agreement on
AI Advisory Board final report (UN, 2024). This report Tariffs and Trade (GATT).78 Ultimately, however, adjudicators
stresses the need for a dispute resolution system that concluded that the measure was not justified because it had
could be facilitated by global forums and explicitly refers not been demonstrated that the aspects of the measure found
to the WTO Dispute Settlement System as an example of to be inconsistent with provisions of the GATT were
dispute resolution “facilitated through global forums”. “necessary” to achieve social inclusion and access to
information (digital divide) within the meaning of Article XX(a).79
While, to date, no disputes on AI measures have
been brought before the WTO Dispute Settlement
System, there have been various disputes related
to aspects of the digital economy. For example, (e) Promoting
disputes have arisen in relation to the tariff treatment of
new technologies and multifunctional products,75 digitally
inclusiveness through
delivered services methods of transmission or delivery,76 special and differential
and whether existing commitments of WTO members cover
new products (e.g., whether terms in specific commitments
treatment and
under the GATS should be interpreted solely according technical assistance
to the meaning they had at the time of entry into force –
i.e., sound recording distribution services).77 Of particular
interest is a WTO dispute which raised issues related WTO agreements recognize the constraints faced by
to the so-called “digital divide”, which, as noted above, is developing economies. They therefore include various
a concern mentioned in various international initiatives special and differential treatment (S&D) provisions
on AI governance. The dispute involved a governmental tohelp them implement WTO rules and participate more
programme which was arguably aimed at “bridging the effectively in international trade. These provisions aim to
digital divide” within that economy. Adjudicators confirmed increase trade opportunities for developing economies and
that, as a general proposition, “the objective of bridging require members to safeguard the interests of developing

80
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Opinion piece Daniel Trefler

Building global chains of trust J. Douglas and Ruth Canada
Research Chair in Competitiveness
During the Industrial Revolution, living standards and Prosperity, Rotman School
in a small group of economies broke free of past of Management; Senior Research
growth trends. Driven on by innovations that Fellow at the Canadian Institute
systematically mobilized science, incomes and for Advanced Studies
public health rapidly improved. It was the epoch of
the “invention of innovation”.

That epoch is about to be repeated: AI represents The WTO is uniquely suited to managing technical
a major re-invention of innovation, positioning disputes. Specifically, the TBT and SPS committees
humanity to revolutionize fields such as healthcare, provide a highly effective forum for technical disputes.
agriculture and material efficiency. However, AI Since 1995, around 56,000 regulatory measures
also introduces unprecedented levels of distrust have been notified to the TBT Committee, with
in the goods and services it creates and powers. only around 830 STCs raised and only 11 disputes
Addressing this distrust is where the WTO can resulting in a panel report. This track record of
play a crucial role, by developing and enforcing soft-law mediation highlights the WTO’s effectiveness
international AI regulations. in technical dispute resolution.

Trust is fundamental in both national and international What makes WTO committees such as TBT and
contexts. Consider the chain of trust involved in SPS committees even more unique is that they bring
treating a child’s fever with antibiotics: from the technical experts together with government officials
doctor’s certification to the drug’s approval by who understand the social dimensions of disputes.
government agencies to enforcement through Thus, technical and social issues are explored
malpractice litigation. This trust ensures the safety simultaneously. In contrast, other standards-setters,
and efficacy of the treatment. such as the 3rd Generation Partnership Project
(3GPP) collaborative project of telecommunications
In international trade, the chain of trust is also fragile. associations, which sets 5G and 6G standards, are
Historically, trade has involved one-sided trust e.g., poorly suited to discussing social values disputes
China exported blue jeans and imported US aircraft. because the discussion can be dominated by certain
Now, with AI-enabled, data-generating products, firms or governments. This does not happen in the
trust must be mutual, not one-sided. TBT and SPS committees.

To address this, we must build an international chain Policymakers are closely focused on global value
of trust. The WTO is well-positioned to contribute chains. They must now become equally attentive to
to this project. The links of the chain separate into the problem of deteriorating global chains of trust.
two broad areas, technical standards and social The WTO has a unique role to play in this.
values. Social values include views on things like
privacy and what constitutes harmful content. Disclaimer
No single international regulatory body can rebuild Opinion pieces are the sole responsibility of their
the many technical and social dimensions of the authors. They do not necessarily reflect the opinions
chain of trust. Multiple approaches are needed. or views of WTO members or the WTO Secretariat.

economies when adopting trade measures. These provisions TBT Committee’s Six Principles,81 in particular Principle 6 on
also grant developing economies flexibilities and longer “Development Dimension”. These WTO instruments play an
implementation periods with respect to their WTO important part in promoting regulatory alignment and stress
obligations and commitments, or are concerned with the the importance of technical assistance to help developing
provision of technical assistance to developing economies. economies overcome their constraints, including in the area
Some WTO S&D provisions apply exclusively to LDCs. of national quality infrastructure (see Chapter 4(b)(iii)). As
Technical assistance and S&D have been stressed in various seen in Chapter 2, investment in AI is unequal across the
WTO “soft law” instruments, such as the TBT Conformity globe, and policy action is largely dominated by developed
Assessment Procedures Guidelines (WTO, 2024b), the two economies. Given the unprecedented opportunities that AI
TBT-related March 2024 Ministerial Declarations80 and the offers to improve productivity and stimulate growth, a lack

81
 CHAPTER 4: WHAT ROLE FOR THE WTO?

of investment in and policy action with regard to AI is likely

to exacerbate further the already significant digital divide.
Additional international financial and technical support is
(f) S
 cholars’ views on the
needed to build the capacity of developing economies in AI possible implications
and enable them to benefit from this technology.
of AI for international
Technical assistance and capacity-building are key trade rules
pillars of the WTO’s work and play a fundamental
role in furthering understanding of the WTO rules and
agreements and of other trade-related topics. Training The WTO provides a framework that can help address
on AI and trade are being integrated into some WTO technical the trade-related aspects of AI governance but the rise
assistance activities; for example, the WTO has incorporated of AI could also have implications for international
the topic of AI and IP into its technical cooperation trade rules. To explore these implications, academics
activities, including two flagship technical assistance events working at the intersection of AI and trade were asked to
(the WIPO-WTO annual colloquium for IP researchers and respond to a survey (see Annex 4).82 This section presents
teachers and the WIPO-WTO annual advanced course on their views based on their responses to the survey and related
topical IP policy issues for government officials). However, the literature. Relaying questions raised by academics and
WTO alone cannot address all of the challenges related experts is important to help better understand the dynamically
to trade, including digital trade and new technologies evolving context in which the WTO agreements operate.
such as AI. Capacity-building, more broadly, is also However, the views expressed do not reflect the positions
frequently provided through various multi-agency and multi- or opinions of WTO members or the Secretariat and are
stakeholder programmes. without prejudice to members’ rights and obligations under
the WTO agreements. The academics’ opinions expressed
Multi-stakeholder programmes like Aid for Trade herein are the sole responsibility of the respective authors.
and the Enhanced Integrated Framework could be
further leveraged to help developing economies AI gives renewed emphasis to some well-known
seize the benefits of AI for trade. Enhancing international issues raised by the increasing digitalization of our
cooperation is critical for making digital trade more inclusive economies. These include issues related to cross-border
(IMF-OECD-UN-WBG-WTO, 2023). The Aid for Trade data flows, data localization, source code, and the blurring
initiative is a WTO-led multi-stakeholder programme launched of the lines between goods and services (see also Box 4.4).83
in 2005 to help developing economies, and in particular
LDCs, to build the trade capacity and infrastructure they need to AI has prompted some academics to wonder about
benefit from trade opening. The initiative can play an important the implications of the technology for international
role, for instance, in supporting the governments of developing trade rules. The unique characteristics of AI, and in
economies in their efforts to enhance connectivity and leverage particular the technology’s capacity to learn, evolve and
technologies like AI for trade by adapting their policies to provide generate outputs autonomously (see Chapter 1), and the
an enabling environment for investment, competition and greater interactivity that this implies, could, these academics
innovation. Some recent Aid for Trade projects have focused argue, pose new challenges for regulators, with ramifications
on sectors such as transport, soft and hard infrastructure, and for trade. For example, some academics wonder whether
agriculture, which already integrate an AI dimension. Such automated legal advice tools, which are increasingly used
projects aim to help beneficiary economies use AI to optimize for a range of tasks by a number of law firms, may comply
transportation or manufacturing processes, or to promote with qualification requirements and how this may relate
sustainable agriculture. Beyond direct support for using AI, Aid to the General Agreement on Trade in Services (GATS).84
for Trade contributes to bolster digital connectivity by fostering Other authors also wonder whether generative AI’s ability
physical and digital infrastructure, both of which are essential to produce output autonomously might not make the
to foster AI deployment. Aid for Trade commitments to the ICT classification or the determination of the origin of certain
sector stood at around US$ 2 billion in 2022. Launched in services more complex.85
2008, the Enhanced Integrated Framework’s institutional and
productive capacity-building projects also help participating Some scholars have stressed the benefits of basing
LDCs to develop digital strategies and skills. In addition to services measures on international standards and
these initiatives, the WTO Secretariat and the World Bank are notifying regulations and standards on such services.
working together on the “Digital Trade for Africa” project. The As already noted, experts have pointed to the fact that
aim of this project is to support efforts by African economies widely differing domestic regulatory approaches may lead
to develop the hard and soft infrastructure necessary to to fragmentation and hamper the opportunities and benefits
harness the opportunities of digital trade. Capacity-building associated with AI. International standards play an important
and the digital divide have also been addressed under the role in promoting regulatory coherence. The GATS, however,
Work Programme on E-commerce (see section 4(a)(ii)). contains limited provisions on standards. In addition, the
Members have shared their own experience with regard to lack of TBT-like disciplines related to technical regulations,
the challenges that they face in building their digital capacity, standards and conformity assessment procedures in the
but also examples of projects and programmes designed GATS may lead members not to notify measures that only
to create a conducive e-commerce environment. apply to AI-enabled services.86

82
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Scholars also note that recent AI developments Agreement’s Annex 3 (“Code of Good Practice for the
may lead members to take a fresh look at the WTO Preparation, Adoption and Application of Standards”) also
reference paper on telecommunications. An expert contains disciplines on domestic standards, which include
has stressed the importance of assessing how the digital not only those adopted by members’ governmental bodies,
transition has impacted competition, for example by making but also those adopted by non-governmental bodies
some markets harder to define, and market dominance more located within a member’s territory. There is an ongoing
difficult to identify. This expert has suggested that disciplines discussion in the TBT Committee on whether “non-
on anti-competitive behaviour in telecommunications, such governmental” standards relate more broadly to “private
as those covered by the reference paper,87 should take standards”, as this is a term not used in the TBT Agreement
account of AI developments, in particular the shift to (WTO, 2021). Given that purely “private” standards (e.g.,
programmable software defined networks and network standards created by industry consortia) may play an
function virtualization – both of which are increasingly important role in AI governance and regulation, an expert
AI-enabled – which allow traffic on telecommunications suggested that consideration could be given to clarifying the
networks to be automatically optimized, and thereby affect meaning of “non-governmental” standards under the TBT
the nature of competition.88 Agreement, including whether or not, and to what extent, this
term may encompass more broadly the concept of “private”
Academics have also suggested that, given the standards. This expert suggested that it could be useful to
pervasive nature of AI and the complexities and discuss how the mechanisms and tools that already exist in
sensitivities of the issues it raises, regulators and the TBT Agreement (i.e., Annex 3: Code of Good Practice
businesses could benefit from notifications of draft for the Preparation, Adoption and Application of Standards)
measures addressing AI, similar to what is done in can be best utilized to ensure that AI standards adopted
the TBT Committee. Regulations related to services are by non-governmental bodies do not result in unnecessary
particularly relevant for AI. According to one expert, one trade restrictions.91 Other experts have suggested that
option could be to introduce a mechanism allowing WTO dialogue with private parties, in particular non-governmental
members to notify draft measures related to AI in the context standard-setting bodies that develop AI standards and
of the GATS Council. Such a mechanism could enhance guidelines, could be strengthened.92
transparency and help to address concerns related to
AI-enabled services.89 According to some experts, current WTO exceptions
may not be sufficient to address the challenges
On the goods side, some academics are of the view raised by AI. These academics note that a fresh look at the
that customs valuation issues and expanding the current language used in current WTO exceptions, which
scope of the Information Technology Agreement is based on a pre-digital age, may be needed to take AI
(ITA) could warrant attention. As noted in Box 4.4, while developments into account.93
the Customs Valuation Agreement and the 1995 Decision
on Valuation of Carrier Media Bearing Software for Data AI also challenges current approaches to IP rights.
Processing Equipment can be useful vehicles to capture As noted in Chapter 3(a)(iv), AI poses challenges to
the value of AI-enabled features of imported goods, the the human-centric approach to IP rights. In addition,
evolutionary nature of AI raises new issues. For example, algorithmic secrecy can prove problematic where there
some experts indicate that the software embedded in is a need to ensure AI’s trustworthiness by investigating
automated vehicles or other AI-enabled devices does not how it has arrived at results (see Section 3(b)(iii)).94
fit squarely with the 1995 Decision. If such software were An expert has noted that balanced IP rights policies need
to fall outside the scope of the Decision, then the question to be put in place worldwide in order to preserve the scope
would be how an electric vehicle embedded with free AI for “freedom to operate” for new entrants. Governments
software that provides for basic self-driving features should and companies trying to join the global knowledge-based
be valued, if it integrated the possibility to upgrade the economy in a world driven by increasingly faster innovation
software for a significant price later on to achieve a much cycles powered by a technology and AI, need access to
higher degree of autonomy. Given the rapid pace of innovation large datasets. This access could be rendered more difficult
and the potential for upgrading hardware that supports AI, where large stocks of data are protected by IP rights. This
as well as for AI’s extensive application in new ICT expert has suggested that certain choices made decades
products, an expert suggested that consideration could ago, when members joined the TRIPS Agreement, may
also be given to expanding the scope of the ITA to further no longer be up to date and could be reviewed against
support AI development and deployment.90 the backdrop of new technologies.95 Meanwhile, some
academics have suggested that consideration could be
A key question raised by academics concerns the given to fostering dialogue in the TRIPS Council to address
role of private parties and non-governmental bodies IP issues raised by AI. Issues that merit particular attention,
in the development of AI-related standards, which in their view, are those related to the use of copyrighted
are key to trustworthy AI. The TBT Agreement contains material to train AI systems, the legal status of AI systems as
various provisions concerning standards. Some provisions creators or inventors, whether AI-generated works are eligible
require WTO members, when appropriate and when for copyright protection, the transparency of algorithms,
possible, to base their TBT measures on existing international and the balance between IP protection and competition,
standards adopted by international bodies. The TBT with adequate IP protection terms.96

83
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Another issue raised in survey responses relates to models.102 Given AI’s fast-evolving and cross-cutting nature
economic rent and competition issues arising from and the significant challenges it is raising, some experts have
AI’s scalability and network effects. As seen in Chapter 2, suggested the need for a “WTO AI and Trade” task force or
AI generates significant economic rents due its scalability working group, or even a dedicated committee.103 Such an
and network effects, leading to market concentration. An expert approach, they reason, would help to overcome “the siloed
has noted that the multilateral rules-based system emerged nature of WTO rules that does not permit addressing AI-related
in a context of low economic rent in a mature, globalized issues adequately”, not least because of the goods-service
industrial economy.97 According to some survey respondents, classification issues,104 and this would, in their view, make it
reviving discussions on competition and technology transfer possible to discuss trade-related issues in one single place
to address the issues raised by an AI-driven rent-rich world in a coordinated manner;105 (ii) facilitate cooperation and
would be worth considering.98 coordination; and (iii) enable more stakeholders to be informed,
get involved and share best practices.106 A recent report by the
Some academics have suggested that AI’s expected World Economic Forum (WEF, 2024a) outlines various possible
disruptive impact on employment may call for new trade areas of work for the WTO, including hosting educational
approaches to mitigating disruptions to labour markets. sessions, conducting a comprehensive assessment of how the
An expert has noted that the WTO Safeguards Agreement, current trading system applies to AI and identifying gaps in
which aims to remedy serious injury, caused by a surge of current rules, encouraging members to present and notify
imports of a specific product, to the domestic industry their AI legislation and regulations, reviewing the implications
producing “like products”, may not capture AI’s potentially of AI for IP rules, developing rules or best practices around
significant impact on tasks performed by humans across all transparency and disclosure with reference to AI use,
economic sectors and industries. This expert argues for the developing guidance on how to facilitate the transparency and
development of a conceptually appropriate approach to verification of AI systems across borders, and discussing the
manage the trade-related impacts of AI adoption that threaten development of AI technical standards.
harm to “tasks across industries, without the pre-condition that
there be a competing ‘industry’ in the importing country”.99 The above-mentioned views suggest that more research
is necessary. Reflections on the implications of AI for trade
Some respondents argue that the current rush rules are still in their early stages. Despite a growing body of
to regulate AI is creating a risk of regulatory literature, more work is needed to fully explore the possible
fragmentation, and it is therefore urgent to find a implications of AI for regulatory frameworks and trade rules.
common ground. In their view, however, AI may not yet Given the speed of AI developments, it is too early to fully grasp
be “treaty-ready” although it may be “discussions-ready”.100 these issues in a definitive manner. It is important to underline
The emerging fragmented regulatory landscape is raising that discussions on the implications of AI for trade rules do
significant concerns, leading to calls for greater international not detract from the rights of WTO members to regulate AI
coherence and multilateral commitments.101 One expert in line with the existing WTO rules. For example, under the
noted that it seemed more likely that a more harmonized GATS, members have the capacity to set non-discriminatory
multilateral approach could be achieved if economies qualification requirements for the supply of services. Rather,
take a balanced and progressive view of regulation, this report is an invitation to explore the potential implications
covering potential regulatory gaps and adopting high-level of AI for international trade, including its rules, with a view to
governance mechanisms rather than overly prescriptive ensuring that we are prepared for the challenges to come.

Box 4.4: systems and models are software

Classification of some digital products: algorithms, although this has not
been discussed in that specific
a long-debated issue in the WTO context in the WTO.109

As AI becomes more and more

Debates about the impact of electronically delivered software. ubiquitous and permeates all
digitalization on how certain products At issue is whether, or under what economic sectors in different and
might be treated under WTO rules circumstances, certain products complex ways, classification issues
are not new. Members have long transmitted via electronic means may resurface. As already discussed,
discussed the classification of certain should be covered under GATS AI also raises issues of IP rights,
digital products in the context of the rules (as services) or GATT rules which are covered by the WTO
WTO Work Programme on Electronic (as goods).108 This question, TRIPS Agreement. All of this may
Commerce, adopted in 1998.107 which remains open in the WTO thus present challenges in terms of
Classification discussions have, context, may be pertinent in the which or how WTO rules apply in
over the years, focused on case of AI, on the basis that AI different contexts.110

84
 CHAPTER 4: WHAT ROLE FOR THE WTO?

Endnotes
1 Thematic Session on regulatory cooperation on “intangible 18 For more information on these issues, see for example National
digital products” organized in the context of the WTO Technical Board of Trade Sweden (2023), Kerry (2024) and Meltzer (2023).
Barriers to Trade (TBT) Committee. See WTO official document
19 See WTO official document G/TBT/M/93 for the minutes of
number G/TBT/GEN/356 (20 July 2023), available at https://
the meeting of 6-7 June 2024, paras. 7.1-7.2.
[Link]/.
20 See WTO official document G/TBT/GEN/385 for the
2 See [Link]
UNECE documents and a brief explanation on the draft guidance
summit-2023-the-bletchley-declaration/the-bletchley-
being discussed.
declaration-by-countries-attending-the-ai-safety-summit-1-2-
november-2023. 21 See [Link]
2506202410_e/sps_2506202410_e.htm.
3 The fundamental role of regulatory transparency was recently
recognised by the panel in EU and certain Member States – Palm 22 TBT Agreement, Articles 2.4, 2.5 (second sentence), 5.4
Oil (Malaysia), Panel Report, para. 7.719. and Annex 3.F. On the presumption under Article 2.5 (second
sentence) see Panel Report, Australia – Tobacco Plain Packaging,
4 Composed of all WTO members, the WTO Technical Barriers
paragraphs 7.254 7.417. The TBT Agreement states that when
to Trade (TBT) Committee is the body responsible for the
an international standard is not an “effective” or “appropriate”
implementation of the TBT Agreement. For more details on the
means for the fulfilment of the legitimate objectives pursued by
functions and work of the Committee, see WTO (2021).
a given regulation, a member is not required to use it as a basis.
5 European Union. See WTO official document number G/TBT/N/ In addition, the TBT Agreement recognizes that developing-
EU/850, available at [Link] economy members should not be expected to use international
standards when these standards are not appropriate in light of
6 [Link] their development, financial and trade needs (Article 12.4).
736&domainId=TBT European Union. See WTO official document
number G/TBT/N/EU/850, available at [Link] 23 See Articles 2.6 and 5.5 and Annex 3.G of the TBT Agreement.
7 Kenya. See WTO official document number G/TBT/N/KEN/1604, 24 See Article 11.2 of the TBT Agreement. See also the 2024 WTO
available at [Link] Ministerial Declaration on “Strengthening Regulatory Cooperation
to Reduce Technical Barriers to Trade” (WT/MIN(24)/35),
8 See Lim (2021). paragraph 5(h) and the 2024 Ministerial Declaration on the
9 See [Link] “precise, effective and operational implementation of special and
differential treatment provisions of the Agreement on the Application
10 See Annex 3 of the TPRM ([Link] of Sanitary and Phytosanitary Measures and the Agreement on
tratop_e/tpr_e/annex3_e.htm). Technical Barriers to Trade” (WT/MIN(24)/36) (available at https://
11 See, e.g., China TPR (2024), Report by the Secretariat [Link]/english/thewto_e/minist_e/mc13_e/documents_e.
(WT/TPR/S/458), paragraphs 21, 34, 3.92, 3.119 and 3.140; htm). See also Principle 6 (“development dimension”), of the
Canada TPR (2024), Report by the Secretariat (WT/TPR/S/455), TBT Committee’s “Six Principles” ([Link]
paragraphs 3.154, 3.227; 3.272, 3.294, and 3.329 3.330; Japan tratop_e/tbt_e/principles_standards_tbt_e.htm).
TPR (2023), Report by the Secretariat (WT/TPR/S/438/Rev.1), 25 See also section 3(c) for a discussion on socio-technical risks.
paragraphs 2.40; 3.134, 3.168; 3.173 and 3.227; and European
Union TPR (2023), Report by the Secretariat (WT/TPR/S/442), 26 See references to ISO/IEC foundational AI standards in Annex
paragraphs 2.51, 3.168 and 3.282. [Link].A. As described by Callegari et al. (2022).”Standards have the
potential to clarify ambiguities and build common understanding
12 WTO official document number IP/C/W/698. Some members around AI risk concepts and terminologies … foundational
expressed their willingness and interest to engage (WTO official standards … are important building blocks in the trustworthy
document number IP/C/M/108/Add.1). AI domain as they lay the groundwork for future assurance
13 For more information see WTO (2022; 2023b; 2023a; 2023d; mechanisms like conformity assessments and certification …
2023c). Given the multistakeholder nature of AI committees, SDOs were
seen to be particularly well placed to achieve consensus around
14 See WTO official document numbers G/TBT/W/788 (16 key concepts such as bias or human oversight. … Nevertheless,
February 2024); G/TBT/W/780/Rev.1 (1 March 2024) and G/ some interviewees urged caution around the role of standards in
TBT/W/789/Rev.1 (23 May 2024). AI ethics. A government official stressed that ‘quite a lot of things
that people are worried about in AI risk is a genuine question of
15 See WT/MIN(22)/32, available at [Link]
ethics or values, where people could completely disagree about
dol2fe/Pages/SS/[Link]?filename=q:/WT/MIN22/32.
the right answer’ and that SDOs are not the right institutions to set
pdf&Open=True.
these values … Instead, standards should enable implementation
16 For more information see: [Link] of agreed-upon values proposed by governments or multilateral
tratop_e/sps_e/sps_2506202410_e/sps_2506202410_e.htm. organisations …. Consequently, for AI risk areas where fundamental
ethical dilemmas persist, standardisation work may face additional
17 See WTO official document number G/SPS/W/361 (22 April
complexities and delays.”
2024). Proposal from Australia under the 6th Review of the SPS
Agreement. Australia observes that digitally enabled solutions are 27 See ISO/IEC Technical Report 24368 (2022): AI – Overview
“increasingly used within the regulatory frameworks that govern of Ethical and Societal Concerns. See also NIST “A Plan for Global
agri food trade”. With respect to AI, specifically, Australia notes Engagement on AI Standards” (available at: [Link]
that “AI platforms also have the potential for assessing compliance [Link]/nistpubs/ai/[Link]). However, others, while
and conformance and implementing real-time follow up and considering that AI standards can address “fundamental rights”,
checking of goods and accompanying documentation.” Australia caution that – in this area at least – this role should be strictly
thus proposed that the SPS Committee put “a strong focus on limited to non-normative issues, e.g., disseminating information
the application of digital technologies [...] as well as the potential and encouraging best practices in processes and measurement
application of artificial intelligence” so as to “ensure that the techniques; standards however “can never attempt to decide on a
benefits and challenges of these technologies can be considered trade off or on a level of acceptability of a given fundamental right
by all Members”. risk” (Gornet and Maxwell, 2024).

85
 CHAPTER 4: WHAT ROLE FOR THE WTO?

28 The EU AIA, for instance, refers to the relevance of mutual allowing it to thrive. Because it signals that WTO members aim to
recognition agreements, that are in line with the WTO TBT keep current customs duties practices on electronic transmissions
Agreement, for facilitating certification procedures of AI systems unchanged, businesses gain the necessary confidence to invest
covered by that regulation. AIA, Preamble, Recital (127). and create jobs. However, some WTO members have expressed
concerns about the lack of clarity in the scope of the moratorium
29 Decision of the Committee on Principles for the Development
and in the definition of electronic transmissions, and the potential
of International Standards, Guides and recommendations with
lost customs revenue. These members have expressed the desire
Relation to Articles 2, 5 and Annex 3 of the Agreement, WTO
to maintain policy space in light of the uncertainty associated with
official document number G/TBT/9, 13 November 2000, para. 20
rapid technological change (IMF-OECD-UN-WBG-WTO, 2023).
and Annex 4.
45 As noted above, AI trustworthiness depends on its ability to
30 See TBT Handbook, pp. 32-33; OECD and WTO (2019, p.
meet stakeholders’ expectations in a “verifiable way”, for example
41-43, 61, 80 & 95-96); and McDaniels et al. (2018, p. 819-821).
via certification against technical specifications in a regulation or
31 For instance, the G7 Trade Ministers’ Digital Trade Principles standard.
make specific reference to the Six Principles as the basis
46 “As AI technologies increasingly underpin the digital services
for developing international standards for information and
we use every day, the importance of the National Quality
communication technology (ICT). See also [Link]
Infrastructure in assuring those AI technologies will be brought
government/news/g7-trade-ministers-digital-trade-principles.
into even sharper focus” (TIC, 2024). As WTO Deputy-Director
32 See, for example, references in UNESCO (2021) to conformity General Jean Marie Paugam said in his opening remarks at the
assessment measures and related instruments. 5th China Quality Conference, “it is clear that digitalisation and
decarbonation have a potential to revolutionize trading patterns and
33 See WTO official document G/TBT/54, Section 2.5 have implications for Quality Infrastructure. Artificial intelligence
(“Acceptance of results”). and other digital products have an immense potential to facilitate
34 See WTO official document G/TBT/W/792 (26 February trade while pushing the frontiers of regulatory cooperation on
2024). More broadly on mutual recognition agreements, see WTO cybersecurity and intangible digital products.” (1 September
Secretariat Note G/TBT/W/42 (28 April 1997). 2023, [Link]
01sep23a_e.pdf). See [Link]
35 Provisions related to data flows, data localization and source events/news/press-release-accredited-tic-sector-key-providing-
code are not included in the stabilized text that was issued on confidence-ethical-ai-development.
26 July 2024 (WTO official document INF/ECOM/87).
47 See TRIPS Agreement, Article 29.1 ([Link]
36 “Services Sectoral Classification List”, WTO official document english/docs_e/legal_e/27-trips_01_e.htm).
[Link]/W/120. The list includes the sector of “computer and
related services”, which refers to category 84 under the Central 48 Under Article 13 of the TRIPS Agreement, the three-step test
Production Classification (Provisional). stipulates that exceptions to copyright protection must only cover
special cases, must not conflict with a normal exploitation of the
37 For the four modes of supply distinguished under the GATS, work, and must not be unreasonably prejudicial to the legitimate
see [Link] interests of the copyright-holder. Similar tests are found in Article
htm#4. 17 for exceptions to trademark rights, and in Article 30 for
38 The 1999 Progress Report on E-commerce adopted by the exceptions to patent rights.
Council for Trade in Services characterized the electronic delivery 49 See TRIPS Agreement, Article 1.1 ([Link]
of services as generally considered to fall within the scope of the english/docs_e/legal_e/27-trips_01_e.htm).
GATS. Dispute settlement cases involving services have, to date,
echoed this line of reasoning. See the Progress Report to the 50 TRIPS Agreement, Article 66.2 ([Link]
General Council, adopted by the Council for Trade in Services docs_e/legal_e/27-trips_01_e.htm).
on 19 July 1999 (WTO official document number S/L/74,
51 TBT Agreement, Preamble, 8th recital.
27 July 1999).
39 In the Services Sectoral Classification List (see [Link] 52 See [Link]
[Link]/english/tratop_e/serv_e/serv_sectors_e.htm), “computer wkgp_trade_transfer_technology_e.htm.
and related services” are composed of five subsectors covering 53 WTO official documents IP/C/R/TTI/CAN/2, 3 and 4; IP/C/R/
different elements of the CPC 84 category: consultancy services TTI/EU/2 and 4; IP/C/R/TTI/CHE/2, 3, and 4. IP/C/R/TTI/USA/2,
related to the installation of computer hardware (CPC 841); 3, and 4, available via [Link]
software implementation services (CPC 842); data processing
services (CPC 843); data base services (CPC 844); other (CPC 54 WTO official document WT/GC/W/443, which requests that a
845+849). Working Group on Trade and Technology Transfer be established,
notes that “the lack of full and faithful implementation of these
40 This does not take into account horizontal limitations that may provisions by developed countries have not allowed developing
affect all sectors within the schedule. GATS mode 4 (movement countries to fully benefit from the growth in international trade”,
of natural persons) commitments tend to refer to horizontal and in document WT/WGTTT/3, members note that “in most
commitments, which are typically “unbound” except for specified cases, however, such provisions contain only ‘best-endeavours’
categories of natural persons. commitments, and are not mandatory rules. The question that
41 In the GATS classification system, the telecommunication arises is to what extent developing countries benefit from these
services sector is composed of 15 subsectors. instruments”. More recently, the African Group noted that, “A
core concern of LDCs has been that while some Members have
42 See the Decision on the Valuation of Carrier Media Bearing made efforts, [...] some of the policies and programmes reported
Software in WTO document G/VAL/5, paragraphs B.2(i) and (ii). by developed countries either barely target or do not at all target
LDCs” (document JOB/TN/CTD/8, JOB/TNC/121). Noting that
43 See WTO document WT/MIN(24)/38.
“Article 66.2 of the TRIPS Agreement places a positive obligation
44 Proponents note that the standstill on customs duties has on developed countries to provide incentives to enterprises and
supported a stable and predictable environment for digital trade, institutions in their territories for the purpose of promoting and

86
 CHAPTER 4: WHAT ROLE FOR THE WTO?

encouraging technology transfer to least developed country 71 There are 18 notifications if the addenda are considered.
Members in order to enable them to create a sound and viable See, e.g., WTO official documents G/TBT/N/KOR/1164, G/
technological base”, LDCs have also “expressed reservations TBT/N/FRA/219, G/TBT/N/DNK/108, G/TBT/N/FRA/203, G/
about the extent to this obligation has been fulfilled” (documents TBT/N/USA/1497, G/TBT/N/TPKM/378 and G/TBT/N/JPN/527.
WT/GC/W/868, G/C/W/825, WT/COMTD/W/270, IP/C/W/ There were no STCs raised on robotics at the time period.
695 and WT/WGTTT/W/33).
72 See, e.g., WTO official documents G/TBT/N/CHN/1742
55 See also WTO official document WT/WGTTT/3. and G/TBT/N/CHN/880. The legitimate public policy objectives
pursued by these measures (as indicated in their notification
56 Under the most-favoured-nation (MFN) principle, WTO
forms) included the prevention of deceptive practices, consumer
members cannot discriminate between their trading partners.
protection and information, quality requirements, harmonization,
This principle is enshrined in several provisions of the WTO
protection of human health or safety, and protection of the
Agreements, such as Article I of the GATT, Article II of the GATS,
environment. The specific problems or challenges they purport to
Articles 2.1 and 5.1.1 of the TBT Agreement and Article 4 of the
address included interoperability, cybersecurity, privacy and data
TRIPS Agreement. Meanwhile, the national treatment principle
regulation, and consumer protection.
provides that imported and locally produced goods shall be
treated equally, at least after the foreign goods have entered the 73 These are: (i) requirements needed for the type approval
market (e.g., Article III of the GATT and Articles 2.1 and 5.1.1 of of the Automated Driving System of fully automated vehicle
the TBT Agreement). The same principle applies to foreign and (STC ID 766); (ii) “On the safety of wheeled vehicles”, including
domestic services (Article XVII of the GATS), and to foreign and as it concerns various advanced autonomous functions (STC ID
local trademarks, copyrights and patents (Article 3 of TRIPS). 687); (iii) the repairability index of various electronic products,
including robot electric lawnmowers (STC ID 657); (iv) criteria
57 See WTO official number G/TBT/GEN/356.
and test procedures for the approval of motor vehicles with
58 See Article 2.3 of the TBT Agreement. respect to their emergency lane keeping system, including with
respect to automated and fully automated vehicles (STC ID 700);
59 Panel Report, EC – Sardines, paras. 7.79-7.82. See
and (v) Internet of Vehicles Cybersecurity Protection Guideline
also EU and certain Member States – Palm Oil (Malaysia),
Rules (STC ID 537).
paragraphs 7.189 (and its footnote 374); 7.567 (and its
footnote 875); and 7.676 (and its footnote 997). See also 74 See Lim (2021) and WTO (2020).
WTO (2020a).
75 EC – Computer Equipment [Link]
60 Article 5.2.7 of the TBT Agreement states that when tratop_e/dispu_e/cases_e/ds62_e.htm and EC – IT Products
product specifications in the technical regulations change, the [Link]
procedures for assessing conformity with them may also need to htm, respectively.
change accordingly.
76 US – Gambling [Link]
61 See also WTO official document G/TBT/GEN/356. dispu_e/cases_e/ds285_e.htm.
62 WTO official document G/IT/25. For a more detailed overview 77 China – Publications and Audiovisual Products [Link]
of all elements of the ITA Committee’s NTM Work Programme, [Link]/english/tratop_e/dispu_e/cases_e/ds363_e.htm.
see WTO (2017).
78 Brazil – Taxation: [Link]
63 See WTO official document G/IT/W/17 and its subsequent dispu_e/cases_e/ds472_e.htm, paragraph 7.583.
revisions, “Draft List of the Types of Conformity Assessment
Procedures for EMC/EMI used by ITA Participants”. 79 Brazil – Taxation: [Link]
dispu_e/cases_e/ds472_e.htm, paragraph 7.622.
64 For example, general-purposes AI models are general by
nature. AI systems, on the other hand, are usually meant to apply 80 i.e., the Ministerial Declaration on “Strengthening regulatory
to specific domains and applications. cooperation to reduce technical barriers to trade” (WT/
MIN(24)/35), paragraph 5(h), and the Ministerial Declaration on
65 See [Link] the “precise, effective and operational implementation of special
imsId=736&domainId=TBT. and differential treatment provisions of the Agreement on the
66 The WTO agreements covered by the Dispute Settlement Application of Sanitary and Phytosanitary Measures and the
Understanding (DSU) are those set out in Appendix 1 to the DSU. Agreement on Technical Barriers to Trade” (WT/MIN(24)/36).

67 For more information, see [Link] 81 See [Link]

tratop_e/dispu_e/dispu_e.htm. standards_tbt_e.htm.

68 While a total of 54 disputes lodged since 1995 have included 82 Responses were received from Susan Aaronson (George
claims of violation of the TBT Agreement, only 11 of these proceeded Washington University), Dan Ciuriak (Centre for International
into actual adjudication by panellists and resulted in panel and/ Governance Innovation), Johannes Fritz (Digital Policy Alert), Olia
or Appellate Body reports. The vast majority of these disputes Kanevskaia (Utrecht University), Kholofelo Kugler (University of
never proceeded beyond consultations, with some ending by Lucerne), Heidi Lund (National Board of Trade Sweden), Petros
virtue of mutually agreed solutions reached by the parties involved. Mavroidis (Columbia Law School), Hildegunn Kyvik Nordås
See WTO (2024a). (Council on Economic Policies (CEP), Örebro University), Eduardo
Paranhos (Associação Brasileira das Empresas de Software)
69 There are 35 notifications if the addenda are considered. and Shin-Yi Peng (National Tsing Hua University).
See, e.g., WTO official documents G/TBT/N/USA/1597, G/
TBT/N/TPKM/399, G/TBT/N/TPKM/400, G/TBT/N/JPN/610, 83 Survey responses by Dan Ciuriak, Johannes Fritz, Kholofelo
G/TBT/N/KOR/776, G/TBT/N/EU/567, G/TBT/N/GBR/36, G/ Kugler, and Shin-Yi Peng. One expert suggested looking into the
TBT/N/TPKM/265, G/TBT/N/USA/2041 and G/TBT/N/GBR/62. classification issue in terms of “durable” products, e.g., music
downloadables, versus “non-durable” products, e.g., streamed
70 There are 30 notifications if the addenda are considered. music (survey response by Dan Ciuriak; see also Ciuriak, 2022).
See, e.g., WTO official documents G/TBT/N/KOR/827, G/
TBT/N/USA/1283, G/TBT/N/JPN/752 and G/TBT/N/ARE/550. 84 See Liu and Lin (2020).

87
 CHAPTER 4: WHAT ROLE FOR THE WTO?

85 See WEF (2024) and survey responses by Hildegunn Kyvik 105 Survey response by Kholofelo Kugler.
Nordås, Kholofelo Kugler and Petros Mavroidis.
106 See Liu and Lin (2020).
86 Survey response by Kholofelo Kugler.
107 Services that are clearly identified as such – e.g., legal services
87 See [Link] or accounting services – and are traded digitally do not pose
tel23_e.htm. classification issues.
88 Survey response by Hildegunn Kyvik Nordås. 108 Classification matters because rules for goods (according to
the GATT, or other specialized WTO agreements addressing trade
89 Survey response by Kholofelo Kugler. in goods) and services (according to the GATS) differ.
90 Survey response by Johannes Fritz. 109 Outside of the WTO, the ISO International Classification
91 Survey response by Olia Kanevskaia. System of standards, which applies to goods, has an entry for
software, and the WIPO Nice Agreement, which provides a
92 Survey response by Dan Ciuriak and Shin-Yi Peng. classification system for goods and services for the registration
93 Survey response by Dan Ciuriak and Shin-Yi Peng. of trademarks, distinguishes between software that can be
downloaded – which is classified as a good under class 9 – and
94 See also survey response by Dan Ciuriak. software that remains on a company’s computer server – which
95 Survey response by Dan Ciuriak. is classified as a service under class 42 (software as a service).
In the UN Provisional Central Product Classification (CPC),
96 Survey response by Dan Ciuriak, Johannes Fritz. from 1991, which is commonly used by WTO members to define
the scope of commitments under the GATS, computer services
97 Survey response by Dan Ciuriak.
comprise various software and computer systems services.
98 Survey response by Susan Aaronson and Dan Ciuriak. See also The more recent version of the CPC (version 2.1) provides
Ciuriak (2024). more detail on computer (or information technology services),
and classifies “software originals” as a distinct sub-category
99 Survey response by Dan Ciuriak. of IT services. The draft CPC version 3, from 2023, clarifies
100 Survey response by Dan Ciuriak. that AI is covered under relevant existing categories, such as
subclasses 83152 “application software provision” and 84392
101 Survey response by Susan Aaronson, Olia Kanevskaia, Heidi “on-line software” ([Link]
Lund and Eduardo Paranhos. CPC/ Documents/4-Accompanying-note-Overview-of-
102 Survey response by Eduardo Paranhos. the-proposed-main-changes-introduced-in-the-revised-
[Link]).
103 Survey responses by Kholofelo Kugler. See also Liu and
Lin (2020). 110 For example, in June 2023, members of the WTO TBT
Committee, on the basis of a proposal by Canada (WTO official
104 Survey response by Kholofelo Kugler. Johannes Fritz also document G/TBT/W/745), held a thematic session on regulatory
notes that “Many AI applications cut across multiple sectors, cooperation on “intangible digital products” (including as they
and core issues like data governance and cybersecurity are relate to AI) under the TBT Agreement which, like the GATT, is an
horizontal in nature. Relying solely on GATS schedules could lead agreement on trade in goods. See [Link]
to fragmentation rather than coherence”. tratop_e/tbt_e/tbt_2006202310_e/tbt_2006202310_e.htm.

88
5 Conclusion
89
CHAPTER 5: CONCLUSION

Conclusion
This report highlights the widespread and transformative to trade and thereby limit the potential of trade to foster
impact that artificial intelligence (AI) is currently having the deployment of trustworthy and safe AI technologies
in many areas, including on international trade, and and the benefits of AI.
discusses the possible future impact of AI in this area. AI has
the potential to reduce trade costs and enhance productivity, As the only rules-based global body dealing with trade
particularly in services sectors that rely on manual policy, the WTO can play a crucial role in limiting regulatory
processes. However, AI also raises important trade-related fragmentation and promoting regulatory coherence. This,
policy questions, in addition to the well-known ethical, in turn, can contribute to the development of AI and
societal and security risks it generates. increase access to it. WTO rules can help to ensure that
AI technologies are beneficial to all economies and
One key challenge lies in addressing the so-called “AI accessible to all by promoting trade-opening in AI-related
divide”, the existing and widening inequality between goods and services. By reducing trade barriers, and thereby
economies with advanced technological infrastructures and fostering a level playing field across economies in terms
those which are less advanced in terms of AI adoption, and of trade, the WTO can encourage the dissemination of AI
between big companies and small businesses. Bridging technologies globally, enabling economies at different stages
this gap is essential to ensure that the benefits of AI of development to access AI innovations. WTO rules can
are equitably distributed across all economies. Another also help to ensure that regulatory interventions are not
challenge concerns the need to access large, accurate more trade-restrictive than necessary and to address
and bias-free datasets to train AI models adequately, and prevent trade tensions and obstacles. However, AI
which must be carefully weighed against the importance may prompt questions about the application of current
of protecting personal data, security and intellectual international trade rules.
property (IP).
The WTO also provides a global framework for cooperation
These questions should be addressed in a coherent and dialogue, within which WTO members can exchange
way across economies, and ways must be found to balance experiences and develop ways to promote trade in
the need to foster global consensus and coherence in AI-enabled products and balance AI risks and opportunities.
AI governance, while respecting diverse cultural and AI governance requires open and inclusive dialogue
societal values. involving all stakeholders, as well as close cooperation
among international organizations. By offering a multilateral
An additional concern is the issue of regulating AI to ensure framework combining predictable and enforceable trade
that it is trustworthy and safe, but without stifling trade. rules with the facilitation of dialogue, the WTO can
This presents a significant challenge for policymakers, meaningfully contribute to the development of a robust AI
given the opacity and autonomous “behaviour” of AI. governance framework and help to create a more coherent,
In addition, while governments across the globe are supportive and inclusive environment for trustworthy and
increasingly taking steps to promote and regulate AI through safe AI.
domestic, regional and international initiatives, the diversity
of these initiatives risks creating a fragmented policy As AI evolves, governments should continue to discuss
landscape. Given the pervasiveness of AI, a coordinated the intersection of AI and trade and its possible implications
global approach involving all stakeholders and international for international trade rules. This report is a first attempt
organizations with a role to play in AI governance is essential to flesh out some of the key implications of AI for trade
to promote policy convergence, as well as to harness and trade rules. It is an invitation to explore these issues
the benefits of AI and mitigate its risks effectively. It is with the aim of ensuring that we fully understand the
important to ensure that differing AI policy approaches opportunities and challenges ahead, and are well-prepared
do not lead to fragmentation, as this could create obstacles to address them.

90
 ANNEX

Annexes
91
ANNEX

92
 ANNEX

Annex 1 Key Supervised versus

unsupervised learning
AI-related terms
• Supervised learning is a machine-learning approach
defined by its use of labelled datasets. These datasets are
This report makes reference to several key concepts designed to train or “supervise” algorithms into classifying
in AI. To facilitate comprehension of these terms, definitions data or predicting outcomes accurately. Using labelled
are provided hereafter. inputs and outputs, the model can measure its accuracy
and learn over time.

General versus narrow AI • Unsupervised learning uses machine-learning

algorithms to analyse and cluster unlabelled datasets.
• General AI or artificial general intelligence (AGI) These algorithms discover hidden patterns in data without
represents a type of AI system that possesses a broad the need for human intervention.
range of capabilities that matches or outmatches humans
(Morris et al., 2024). True AGI systems do not yet exist. The In certain AI models, the distinction between supervised
concept of AGI remains a visionary goal, but the rapid pace and unsupervised learning is more nuanced than in others.
of development of AI hints at the possibilities and potential For instance, in reinforcement learning, the machine is given
directions AGI might take. only a numerical performance score as guidance, and in
weak or semi-supervision models, a small portion of the data
• Narrow AI: Narrow AI refers to a type of AI system that are tagged.
is designed to address specific tasks or solve particular
problems. Unlike AGI, which aims for broad capabilities,
narrow AI focuses on defined tasks and exhibits expertise Other terms
within a limited domain. Narrow AI systems are tailored to
excel in specific applications or problem domains. • Foundation models are large-scale, pre-trained models
that serve as the basis or foundation for developing more
specialized AI applications or models. These foundation
AI technologies models are typically trained on vast amounts of data using
techniques such as unsupervised learning. Developers can
• Machine learning is a subset of artificial intelligence fine-tune these pre-trained foundation models on specific
(AI) that focuses on the development of algorithms and datasets or tasks to create more specialized AI models
statistical models that enable computers to perform tasks tailored to particular applications or domains.
without being explicitly programmed to do so. In other words,
machine-learning algorithms learn from data, identify patterns • Source code refers to the human-readable instructions
and make decisions or predictions based on those data. written by programmers to define the behaviour, algorithms
and models used in AI systems.
• A neural network is a computational model inspired by
the structure and function of the human brain, composed • Artificial Intelligence of Things (AIoT) refers to the
of interconnected nodes, or artificial neurons, organized in integration of artificial intelligence (AI) technologies with
layers. Through a process called training, neural networks Internet of Things (IoT) devices and systems. AIoT combines
learn from examples by adjusting the weights of connections the capabilities of AI algorithms with the vast amounts of
to minimize the difference between predicted and actual data generated by IoT devices to create intelligent and
outputs, thereby enabling them to recognize patterns, autonomous systems.
make predictions and perform complex tasks across a wide
range of domains. • Intelligent automation combines AI technologies, such
as machine learning, computer vision, natural language
• Deep learning is a subset of machine learning that processing and robotics process automation, to automate
involves training artificial neural networks with many layers and optimize processes, tasks and workflows in various
of processing units, or neurons, to learn representations domains and industries.
of data. The term “deep” refers to the depth of the neural
networks, which typically consist of multiple hidden layers
between the input and output layers.

• Large language models (LLM) are advanced AI systems

that are trained on massive amounts of text data to
understand and generate human-like language. These
models are characterized by their vast size, often containing
hundreds of millions to billions of parameters, which
enables them to capture intricate patterns and nuances
in language.

93
ANNEX

or additional AI software, based both on AI and human

Annex 2 Technical judgement. Summing over all tasks for each occupation, an AI
Appendix on the exposure indicator at the occupation level can be calculated.
The AI exposure indicator is then aggregated to the sectoral
simulation scenarios level using employment data, differentiating by skill level (high,
medium and low).4

Productivity shocks To distinguish the global synergy scenarios from the tech
divergence ones, it is assumed that the pattern of productivity
The changes projected as a result of AI in labour productivity increases is on aggregate (economy-wise) reversed between
differ according to skills and sectors,1 distinguishing between middle-skilled and high-skilled labour.5 In the global synergy
high-skilled, medium-skilled and low-skilled labour. scenarios, productivity increase is higher for middle-skilled
workers than high-skilled worker, whereas in the tech
The size of the productivity shock, or changes to productivity, divergence scenarios, higher-skilled workers see a higher
in the optimistic scenarios is based on a study conducted by increase in productivity. However, the sectoral distribution of
Goldman Sachs (2023). The study projects that AI will increase AI exposure is kept, as well as the relative gap between high-
total factor productivity in the United States by 1.5 percentage skilled and middle-skilled workers across sectors. Hence,
points annually for 10 years, starting in 2027, 10 years sectors with larger AI exposure of middle-skilled relative
after AI started to transform the technology industry. Since to high-skilled workers will continue to exhibit a larger gap
the productivity shock will be phased in over 14 years (2027- compared to sectors where the gap is smaller.
40) in the simulation conducted for this report, this implies
an approximate shock of 1.06 percentage points per year. The projections for the level and variation in productivity shocks
have been developed for the United States and applied for
The size of the productivity shock in the cautious scenarios is other economies. However, since productivity shocks vary
partially based on Acemoglu (2024), who projects that total according to skill and sector, and other regions do not have
factor productivity will go up by 0.66 percentage points in 10 the same industrial and skill structures as the United States,
years as a result of AI. However, Acemoglu (2024) follows productivity gains will differ across economies. Regions with a
Svanberg et al. (2024) in assuming that only 23 per cent greater proportion of middle-skilled and high-skilled workers,
of AI projects can be profitably implemented. Since a long- and with a larger presence in sectors with the highest projected
term perspective has been employed here, this profitability productivity shocks, will achieve higher average productivity
scaling-down is not applied, which thus leads to a productivity growth. Also, the degree of convergence between middle and
shock of 0.2 percentage points per year. high-skilled workers may differ.

In the global synergy scenarios, the productivity shocks are

applied uniformly across economies. In the tech divergence Trade cost shocks
scenarios, the productivity shocks are applied taking into
account variation in the International Monetary Fund (IMF)’s As discussed in Chapter 2, AI is projected to impact trade costs
AI Preparedness Index2 across economies. The IMF’s AI through three channels: diminished compliance costs, reduced
Preparedness Index contains indicators of digital connectivity, language barriers and improved logistics. To gauge the size of
skills, innovation capacity and regulations. A replicated AI the impact, insights from the literature concerning the impact
preparedness index was used for this report, based on IMF of AI have been considered. According to these estimations,
methodology (Cazzaniga et al., 2024). This implies that low- reductions in trade costs can occur per the three channels:
income economies would benefit less from productivity
increases, because they score lower on this index. To scale the • A reduction in the costs of compliance with regulations.
productivity increase according to economies’ AI preparedness, To determine the associated trade cost reduction, a measure
the productivity shocks calculated for the United States are for compliance costs has been included in a regression of
multiplied by the AI Preparedness Index of each region relative inferred trade costs (employing the Head and Ries (2001)
to the United States. formula) on trade cost proxies. The compliance cost measure
employed is the World Bank Group’s “Doing Business”
In both the global synergy and tech divergence scenarios, indicator, Documentary compliance to export, defined as the
productivity shocks vary by skill level and across economies. hours needed to comply with all documentary requirements
The variation by skill level is based on literature identifying to export. This generates the ad valorem equivalent trade
which tasks can be automated through AI in the O*NET costs of the documentary compliance. In the counterfactual
catalogue3 of tasks for each occupation. O*NET describes scenarios, it is assumed that these trade costs fall by 70
each occupation in terms of various domains, such as work per cent. This value is based on a case study of DHL’s
activities or tasks that need to be completed. Both the global experience: when using AI-based intelligent document
synergy and tech divergence scenarios are based on the AI processing to prepare the necessary documentation for
exposure indicator by Eloundou et al. (2023), who use the international shipments, DHL observed an efficiency gain
tasks attributed to occupations by O*NET. They calculate of 70 per cent.6 Since costs of documentary compliance
whether the time spent on completing a task can be reduced are proportional to the time needed to comply, this implies
by 50 per cent or more through the application of ChatGPT that trade costs can also fall by 70 per cent.

94
 ANNEX

• A reduction in the costs associated with language barriers • A reduction in logistics costs, since AI is expected to reduce
in international trade, since AI will facilitate translation of the costs associated with logistical planning. To determine
written and spoken communication. To determine the impact the size of the effect, we use the “timeliness” component
on trade costs, the ad valorem equivalent trade cost of the World Bank’s Logistics Performance Index (LPI).7
associated with a dummy for common official language, as The LPI reflects the frequency with which shipments are
introduced by Melitz and Toubal (2014), was employed. In delivered within scheduled or expected delivery times
the regression, spoken common language was controlled (World Bank, 2023). To capture the potential impact of AI
for, as this captures the influence of common language on on logistics costs, the difference between the maximum
trade and trade costs through, for example, ease in informal possible value for this indicator (five) and the actual indicator
communication and building trust in networks. Therefore, it was calculated and included in the regression of inferred
is assumed that trade costs associated with a different trade costs on trade cost proxies. In the counterfactual
official language completely disappear, implying a global a scenarios, the associated trade cost reduction is 50 per
verage ad valorem equivalent trade cost reduction of cent, i.e. it is assumed that AI will improve the timeliness
2.12 per cent. This is close to the projected trade of shipments by decreasing delays by half. As developing
cost reduction of AI through improved machine translation economies and LDCs tend to have a higher frequency of
on eBay in a study by Brynjolfsson et al. (2019), in which delays, such an improvement will contribute to convergence
ad valorem equivalent implied by the projected trade effect by reducing the gap in the frequency of delays between
effect was 2.2 per cent. these economies and developed economies.

Figure A.1: Cumulative trade cost ad valorem equivalents (%) averaged by importers (2017-40)

Global Synergy
0

-5
Cumulative AVE

-10

-15

-20

Tech Divergence
0

-5
Cumulative AVE

-10

-15

-20
asl aus bra can chn e27 eft gbr idn ind jpn kor lac mex min oas row rus sea ssl sso tur usa zaf

Compliance costs Language barriers Logistics costs

Note: Figure A.1 demonstrates projected cumulative ad valorem equivalents of trade cost reductions in the global synergy and
tech divergence scenarios by means of compliance, language and logistics. The values are not additive. See Table A.1 for a list of
abbreviations for region names.
Source: Simulations using the WTO Global Trade Model

95
ANNEX

These trade cost shocks are not assumed to vary between

the optimistic and cautious scenarios. However, they vary
Figure A.2: Cumulative global GDP growth
according to whether the global synergy scenarios or the
rate (2023-40)
tech divergence scenarios are considered. In the global
synergy scenarios, it is assumed that all regions have the same
(Difference to baseline, percentage points)
means to implement AI, and that trade cost shocks are
therefore identical. In the tech divergence scenarios, the IMF’s
AI Preparedness Index has been bilateralized and used to
scale the trade cost shocks to account for differences in 12
regions’ capacity to use AI technologies. It is assumed that 10.85
trade costs decrease between 2018 and 2040 because
some the trends described have already started, such as the 10

reduced costs associated with language barriers referred to in

8.23
the study of Brynjolfsson et al. (2019).
8

Figure A.1 displays the projected cumulative trade cost

reduction under the global synergy scenario (without 6
5.81
considering AI preparedness of economies) and the tech
divergence scenario (considering AI preparedness) over the 4.31
period from 2017 to 2040.8 The figure clearly shows that the 4
largest projected trade cost reductions will be in LDCs and
developing economies. This is because these regions have
2
the greatest potential to leverage AI to lower compliance
costs, logistics costs and costs related to language barriers.
The largest projected trade cost reduction is expected to 0
come from decreased logistics costs for LDCs and developing Tech Global Tech Global
Divergence Synergy Divergence Synergy
economies, while diminishing language barriers play a major Cautious Cautious Optimistic Optimistic
role for some developed regions, such as the European
Union and the European Free Trade Association (EFTA).
Trade costs Productivity growth

Figure A.2 shows the projected changes in real GDP Note: This figure demonstrates the impact of AI
between 2023 and 2040. In the optimistic global on projected cumulative global GDP growth
synergy scenario, global real GDP is expected to grow (as percentages) between 2023 and 2040 in four
by 11 percentage points compared to the baseline over scenarios. The values represent the deviation from
the baseline scenario.
the period, highlighting the impact of AI in boosting
Source: Simulations using the WTO Global
the global economy. The results mirror those for trade, Trade Model
though labour productivity plays a more significant role in
driving GDP growth.

96
 ANNEX

Table A.1: Abbreviations

Regions Sectors
AUS Australia AGR Agriculture
OAS Other Asian countries OIL Oil
CHN China ONR Other natural resources
JPN Japan PRF Processed food
KOR Republic of Korea TWL Textiles, wearing apparel and leather
SEA ASEAN P_C Petroleum, coal products
ASL Asian LDCs CHE Chemicals and petrochemicals
IND India PRP Pharmaceuticals, rubber and plastic products
IDN Indonesia OTG Other goods
CAN Canada MET Metals
USA USA EEQ Electronic equipment
MEX Mexico ELE Computer, electronic and optical products
BRA Brazil OMF Other machinery
LAC Latin America MVT Motor vehicles
E27 EU-27 OTN Transport equipment not elsewhere classified (n.e.c.)
GBR United Kingdom UTC Utilities and construction
EFT EFTA countries TRD Trade
ROW Rest of World TRP Transport
RUS Russian Federation WIS Accommodation, food and services activities
MIN Middle East and North Africa WHS Warehousing and support activities
TUR Türkiye CMN Communication
SSA Sub-Saharan Africa OBS Business Services
SSL Sub-Saharan LDCs INS Insurance
ZAF South Africa FIN Financial Services
OTS Other Services
EDH Education and human health
ROS Recreational and other services

Endnotes
1 In the Global Trade Analysis Project (GTAP) Data Base (https:// 5 It is worth noticing that such an approach does not mean that
[Link]/databases/), “off_pros” are mapped all sectors will have a larger productivity shock in middle-skill
to “high-skilled”, “tech_aspros”, “clerk” and “service_shop” are occupations than in high-skill occupations.
mapped to “medium-skilled”, and “ag_othlowsk” are mapped to
6 See [Link]
“low-skilled”.
dhl-group-increases-efficiency-by-70-with-rpa-and-abbyy-idp/.
2 See [Link]
7 See [Link]
3 A survey-based database of information on jobs and occupations
8 We assume that trade cost reductions begin earlier, in 2017,
([Link]
compared to the increase in productivity starting in 2023, as
4 The skill level of each occupation is based on the International empirical evidence suggests that AI has already contributed to
Standard Classification of Occupations (ISCO-08) major groups lowering logistical and translation costs.
score per occupation (in which 1-3 are high-skill; 4-8 are medium-
skill and 9 are low-skill). This classification is mapped to US
Standard Occupational Classification (SOC2018) occupations.

97
ANNEX

terminology and definitions may be a particularly important

Annex 3 Key Policy element for ensuring coherence and interoperability across
economies’ regulatory interventions in this area (Meltzer,
and Standard-Related 2023). The OECD Principles contain various AI definitions,
International Initiatives of which two – “AI system”1 and “AI system lifecycle”2 – are
key for the implementation of any national AI strategy, policy
in the Area of AI or regulation. As explained in this report, regulatory
fragmentation itself can represent an important trade barrier,
in particular for developing economies and MSMEs.

I. Policy-related initiatives In February 2020, the OECD launched the AI Policy

in the area of AI Observatory, an inclusive hub for public policy on AI. The
AI Policy Observatory aims to help economies encourage
and monitor the responsible development of trustworthy
AI systems. It is intended to facilitate dialogue and provide
A. Organisation for Economic multidisciplinary, evidence-based policy analysis and data on
Co-operation and areas impacted by AI. For governments, it is also intended to
Development (OECD) serve as a centre for policy evidence collection and debate,
with support from strong partnerships with a wide spectrum
In May 2019, the OECD Council adopted the OECD of external actors.3
AI Principles (OECD, 2019a), considered to be the first
intergovernmental policy instrument in this area (Ebers and
Navas Navarro, 2020; Morley et al., 2020). The document B. G20
includes the following five high level values based principles
for responsible AI stewardship: In June 2019, G20 economies committed to a human-centred
AI, and, to this end, adopted non-binding AI Principles.4
1) “inclusive growth, sustainable development and well-being”; The G20 AI Principles were drawn from the OECD AI
2) “human-centred values and fairness”; Principles described above, which include five high-level
3) “transparency and explainability”; principles for responsible stewardship of trustworthy AI and
4) “robustness, security and safety”; and five recommendations for national policies and international
5) “accountability”. cooperation in the area of AI.5

The document also provides five policy recommendations Since 2019, the G20 has been implementing aspects of the
intended to guide both national government policies and AI Principles.6 For instance, to foster knowledge on existing
international cooperation, to be undertaken in a manner approaches and practices, the G20 launched the “Examples
consistent with the five AI principles, namely: of National Policies to Advance the G20 AI Principles”,7 and
the “Policy Examples on How to Enhance the Adoption of
1) investing in AI research and development; AI by MSMEs and Start-ups”.8 AI policy issues were a key
2) fostering a digital ecosystem for AI; issue at the G20 Summit in Rio de Janeiro in November 2024,
3) shaping an enabling policy environment for AI; which focused on the use of AI for sustainable development.
4) building human capacity and preparing for labour market
transformation; and
5) international cooperation for trustworthy AI. C. Council of Europe
The OECD AI Principles have been recognized by economies In 2021, the Council of Europe’s Committee on AI was tasked
beyond the 38 OECD members. Notably, the G20 AI to prepare a legally-binding international instrument on the
principles (see below) – which are mostly based on the OECD development, design and application of AI, based on the
principles – were adopted by various non-OECD economies, Council’s standards on human rights, democracy and the rule
including Brazil, China and India. of law.9 These negotiations concluded in 17 May 2024 with the
adoption of a “Framework Convention on Artificial Intelligence,
The policy recommendations of the OECD Principles Human Rights, Democracy and the Rule of Law”, the first
suggest elements that, directly or indirectly, may relate to trade binding international instrument on AI policy. The Framework
and WTO issues. For instance, they refer to the fact that Convention aims to ensure that activities within the lifecycle of
national policies and international cooperation need to include AI systems are fully consistent with human rights, democracy
the preparation and use of regulatory instruments, such as and the rule of law, while being conducive to technological
technical standards, conformity assessment (certification and progress and innovation (Council of Europe, 2024). It sets
verification) and international standards for interoperable and out several fundamental principles related to activities within
trustworthy AI. the AI systems lifecycle, such as human dignity and individual
autonomy, equality and non-discrimination, transparency and
The OECD Principles also propose a common understanding oversight, respect for privacy and personal data protection,
of certain key AI terms. Global agreement over key AI accountability and responsibility, reliability, and safe innovation.

98
 ANNEX

The Framework Convention also sets requirements to ensure Developing Advanced AI Systems”12 and the “International
the availability of remedies, procedural rights and safeguards, Code of Conduct for Organizations Developing Advanced AI
as well as requirements for risk and impact management. Systems”.13 These documents, which are based on the OECD
It states that its membership is open not only to the members AI Principles and take into account recent developments
states of the Council of Europe, but also to non-members, in advanced AI systems, aim to promote the safety and
under certain conditions. trustworthiness of AI systems by providing guidance, in the
form of principles and actions, for organizations developing
and using the most advanced AI systems.
D. United Nations Educational,
Scientific and Cultural Several principles of the “International Guiding Principles for
Organization (UNESCO) Organizations Developing Advanced AI Systems”14
are particularly relevant for trade. They include: taking
In November 2021, UNESCO’s 193 members adopted appropriate measures to identify, evaluate, and mitigate
the first-ever global policy instrument on AI ethics – a risks across the AI lifecycle; investing in and implementing
non-binding “Recommendation on the Ethics of Artificial robust security controls, including physical security,
Intelligence” (UNESCO, 2021). The Recommendation cybersecurity and insider threat safeguards across the AI
is designed to guide the responsible development and lifecycle; advancing the development of and, where
application of AI technologies, ensuring that they are aligned appropriate, the adoption of international technical standards;
with human rights and ethical standards.10 It provides a set and implementing appropriate data input measures and
of ten core principles, to be followed by all actors in the AI protections for personal data and IP.
system lifecycle, that encapsulate a human rights approach
to AI, emphasizing the importance of safety, security, The actions proposed in the “International Code of Conduct for
privacy, transparency, responsibility, accountability and Organizations Developing Advanced AI Systems”15 include:
non-discrimination. It also lays out the following values: measures to identify, evaluate and mitigate risks across the AI
“respect, protection and promotion of human rights and lifecycle (such as employing diverse internal and independent
fundamental freedoms and human dignity”; “environment and external testing measures and implementing appropriate
ecosystem flourishing”; “ensuring diversity and inclusiveness”; mitigation to address identified risks and vulnerabilities);
and “living in peaceful, just and interconnected societies”. implementing robust security controls, including cybersecurity
policies across the AI lifecycle; advancing the development
The Recommendation also sets out eleven key areas for of and, where appropriate, adoption of international technical
policy actions which call for the development of international standards; and implementing appropriate data input measures
standards to ensure the safety and security of AI systems, and protections for personal data and IP. In July 2024, the
achieving accountability and responsibility for the content OECD announced a pilot phase to monitor the application
and outcomes of AI systems, and fostering research at of the G7 Hiroshima Process International Code of Conduct
the intersection between AI and intellectual property (IP). for Organisations Developing Advanced AI Systems.16
At their October 2024 meeting, Digital and Tech Ministers
To assist its members in implementing the Recommendation, announced that they would continue to work to develop
UNESCO has developed the “Readiness Assessment the Reporting Framework with the aim to advance it by the
Methodology” (UNESCO, 2023a), a tool aimed at evaluating end of the year, in collaboration with the OECD and the
preparedness for the ethical deployment of AI. participating organizations.17

Like the other AI initiatives described in this section, the In the G7 Verona and Trento Ministerial Declaration, adopted
Recommendation contains various elements that relate to in March 2024, beyond advancing these actions under
WTO agreements and issues. For instance, it stresses the Hiroshima AI Process, G7 economies also expressed their
need to develop international standards (see Chapter 4(a) desire to participate in the discussions initiated by the
(iii) on TBT) as tools to support AI policies, regulations and Brazilian G20 Presidency on the specific issue of “AI for
standards adopted in furtherance of the principles and policy sustainable development”.
actions proposed by the Recommendation. It also refers to
the importance of discussing the intersection between AI Relatedly, the G7 has played an important role in developing
and IP (see Chapter 4(b)(iv) on trade-related aspects of and operationalizing the notion of “Data Free Flow with Trust”
intellectual property rights (TRIPS). (DFFT) (Meltzer, 2023).18 For example, the G7 Digital Trade
Principles provide that “data should be able to flow freely
across borders with trust”, and call for unjustified obstacles
E. G7 to cross-border data flows to be addressed, on the one hand,
and for privacy, data protection, the protection of IP rights,
In May 2023, G7 leaders established the “Hiroshima Process and security, on the other.19 In April 2023, the G7 agreed
on Generative AI” with the aim of promoting safe, secure to establish the “Institutional Arrangement for Partnership” to
and trustworthy AI.11 In this context, in December 2023, operationalize the DFFT concept through principles-based,
the G7 Digital and Tech Ministers agreed on the Hiroshima solutions-oriented, evidence-based, multi-stakeholder and
AI Process “Comprehensive Policy Framework”, which cross-sectoral cooperation (see also Chapter 3(b)(i)) for a
includes “International Guiding Principles for Organizations discussion of cross-border data flows).20

99
 ANNEX

The G7 also recognizes the importance of interoperability society and representatives from the international
between tools for trustworthy AI (such as regulatory scientific panel.
and non-regulatory frameworks and technical standards).21 • the creation of an AI capacity development network to
In this context, the G7 has developed an “Action Plan link up a set of collaborating, United Nations-affiliated
for Promoting Global Interoperability between Tools for capacity development centres making available expertise,
Trustworthy AI” in which G7 economies have pledged to compute and AI training data to key actors.
raise awareness of international AI technical standards • the creation of a global fund for AI managed by an
development efforts, build capacity among stakeholders independent governance structure.
on ways to actively participate in such processes, and • the creation of a global AI data framework that would
encourage adoption of international AI standards as tools outline data-related definitions and principles for global
for advancing trustworthy AI.22 governance of AI training data, establish common
standards around AI training data provenance and use,
and institute market-shaping data stewardship and
F. United Nations (UN) exchange mechanisms for enabling flourishing local AI
AI Advisory Body ecosystems globally;
• the creation of an AI office within the Secretariat, reporting
In October 2023, the UN Secretary-General formed a to the Secretary-General.
high‑level AI Advisory Body, composed of experts from
government, industry, academia and civil society, to develop
a set of recommendations on the international governance G. Bletchley process
of AI.23 The Final Report of the UN AIAB was published in
September 2024 (UN, 2024). In November 2023, the United Kingdom hosted the AI
Safety Summit, at which 28 economies and the European
Rather than proposing any single model for AI governance, Union agreed on the “Bletchley Declaration” on AI Safety.
the UN AI Report outlines five guiding principles for the The Summit brought together various governmental and
creation of new AI governance institutions. non-governmental stakeholders to discuss how to mitigate
the risks posed by AI through internationally coordinated
The five guiding principles concern: action (UK Government, 2023). The Bletchley Declaration
recognizes the urgent need to understand and collectively
1) inclusivity (AI “should be governed inclusively, by and for the manage potential risks through a new joint global effort
benefit of all”); to ensure AI is developed and deployed in a safe,
2) public interest (“AI must be governed in the public interest”); responsible way for the benefit of the global community.
3) “data governance” (“AI governance should be built in step It agrees to focus cooperation on identifying common AI
with data governance and the promotion of data commons”); safety risks and building a shared scientific and evidence
4) universality (“AI governance must be universal, networked based understanding of these risks, and building respective
and rooted in adaptative multi-stakeholder collaboration”); risk-based policies across countries to ensure safety in
and light of such risks, collaborating as appropriate while
5) “international law” (“AI governance should be anchored in recognising that approaches may differ based on national
the UN Charter, International Human Rights Law, and other circumstances and applicable legal frameworks. Indeed,
agreed international commitments such as the Sustainable international cooperation is a key tenet stressed throughout
Development Goals”). the text of the Bletchley Declaration.

The UN AI Report identifies three main governance As part of the commitment to international cooperation
gaps - representation gaps, coordination gaps, and and building a shared scientific and evidence-based
implementation gaps – and formulates recommendations understanding of certain AI risks under the Bletchley
to “advance a holistic vision for a globally networked, agile Declaration, the attending economies also agreed to
and flexible approach to governing AI for humanity, support the development of an independent and inclusive
encompassing common understanding, common ground “State of the Science” Report on “frontier AI”.24 The interim
and common benefits to enhance representation, enable version of this report was published in May 2024.25
coordination and strengthen implementation”. Specific The final version of the Report is expected to be published
recommendations include: ahead of the next AI summit, scheduled for February 2025
(UK Government, 2023b).
• the creation of an independent international scientific panel
on AI, made up of diverse multidisciplinary experts in the
field serving in their personal capacity on a voluntary basis. H. UN General Assembly
• the launch of a twice-yearly intergovernmental and
multi-stakeholder policy dialogue on AI governance on In March 2024, the UN General Assembly unanimously
the margins of existing meetings at the United Nations. adopted a non-binding resolution on seizing the opportunities
• the creation of an AI standards exchange, bringing together of safe, secure and trustworthy AI systems for sustainable
representatives from national and international standard- development (UN AI Resolution).26 Although certain UN
development organizations, technology companies, civil specialized agencies (including UNESCO, as noted above)

100
 ANNEX

have adopted AI-related instruments, this resolution is the questions likely to face IP policymakers as AI increases in
first adopted on a UN-wide basis. importance. The key points generated from these debates
were compiled in a WIPO Secretariat “Issues Paper on
The UN AI Resolution establishes a vision that AI systems27 Intellectual Property Policy and Artificial Intelligence”
should be human-centric, reliable, explainable, ethical (WIPO, 2020). The issues identified in the paper included
and inclusive, as well as oriented toward sustainable patents, copyright and related rights, data, designs,
development. It recognizes the “rapid acceleration” of the trademarks, trade secrets, the technology gap and capacity-
design, development, deployment and use of AI systems and building, and accountability for IP administrative decisions.
their potential to contribute to “accelerating the achievement”
of the UN Sustainable Development Goals (SDGs). WIPO has also developed an “AI and IP Clearing House”
Consequently, it stresses the “urgency of achieving global and an “IP policy toolkit”. WIPO’s AI and IP Clearing
consensus” on safe, secure and trustworthy AI systems such House is a searchable database that “continuously
as by promoting the following actions: collates and publishes the main government instruments
of relevance to AI and IP with the aid of the Member
1) developing regulatory and governance approaches and States”.29 The “IP policy toolkit” (WIPO, 2024) is intended to
frameworks; allow policymakers to engage on “how to best shape
2) promoting internationally interoperable identification, their AI innovation ecosystem and to structure their
classification, evaluation, testing, prevention and mitigation future work with a firm understanding of the current state
of risks of AI systems; of knowledge”.
3) developing mechanisms of risk monitoring and
management and for securing data across the lifecycle of
AI systems; J. International Telecommunication
4) developing internationally interoperable technical tools, Union (ITU) (AI for Good platform)
standards or practices;
5) respecting IP rights, including copyright protected content; The ITU, in partnership with 40 UN bodies, has convened
6) safeguarding privacy and the protection of personal the “AI for Good” platform, the goal of which is to identify
data when testing and evaluating systems; practical applications of AI to advance the UN SDGs.
7) promoting transparency, predictability, reliability, “AI for Good” consists of a year-round online programme
understandability and human oversight of AI systems; and and an annual “AI for Good” Global Summit.30 The ITU has
8) sharing best practices on, and promoting international also launched a global AI Repository to identify AI-related
cooperation in, “data governance” for greater consistency projects, research and other initiatives that can accelerate
and interoperability, where feasible, of approaches for progress towards the SDGs.
advancing trusted “cross-border data flows” for safe,
secure and trustworthy AI systems.

More broadly, the resolution recognizes that the AI

II. Standardization in the
governance is still an “evolving area”. As such, it stresses the area of AI
need for “continued discussions” on possible governance
approaches that are “appropriate, based on international
law, interoperable, agile, adaptable, inclusive, responsive There are significant efforts underway on developing
to the different needs and capacities of developed and international standards on AI. Such activities are taking
developing countries alike and for the benefit of all”. In place inter alia in the Joint Technical Committee of the
this respect, the resolution calls for AI “regulatory and International Organization for Standardization (ISO) and the
governance approaches” to be developed based on International Electrotechnical Commission (IEC), the ITU and
inputs from many stakeholders, i.e., the private sector, the Institute of Electrical and Electronics Engineers (IEEE)
international and regional organizations, civil society, the (Meltzer, 2023).
media, academia and research institutions, technical
communities and individuals.
A. ISO/IEC
I. World Intellectual Property In 2018, the ISO/IEC Joint Technical Committee (JTC) 131
Organization (WIPO) established a subcommittee to work exclusively on AI
standardization. As the focal point of standardization on AI
From 2019 to 2020, WIPO, as the UN specialized within the ISO and IEC, the committee looks at the entire
organization dedicated to IP issues globally, held a series AI ecosystem and provides guidance to ISO and IEC
of “Conversations” on the impact of AI on IP policy.28 committees developing AI applications. Its work programme
As discussed in Box 3.2 (Chapter 3) of this report, comprises standardization in the areas of foundational
AI technologies can raise important questions concerning AI standards, data standards related to AI, big data and
the creation and protection of IP rights. In the WIPO analytics, AI trustworthiness, governance implications of
“Conversations”, governments and other stakeholders AI, testing of AI systems, and ethical and societal concerns
debated and submitted inputs about the most pressing (ISO/IEC, 2024).

101
 ANNEX

The ISO/IEC Joint Technical Committee and subcommittee telecommunications, information technology and power-
have already published 25 standards on AI32 and are generation products and services.37 Standards related
currently working on developing another 31 AI standards.33 to AI developed by the IEEE include standards for
Among the standards already published are standards addressing ethical concerns during system design, for
on concepts and terminology, risk management and safety transparency of autonomous systems, and for algorithmic
of AI systems. The ISO/IEC Joint Technical Committee bias considerations.
and subcommittee also published a technical report in
2022, containing an extensive overview on the issue of
“ethical and societal concerns” related to AI D. United Nations Economic
governance.34 Standards still under development cover Commission for Europe (UNECE)
a wide range of new topics such as: “requirements
for bodies providing audit and certification of artificial In February 2023, UNECE launched a new project aiming
intelligence management systems”, “guidance on at developing new guidance on digital product regulation
addressing societal concerns and ethical considerations”, focused on regulatory compliance of “products with embedded
“environmental sustainability aspects of AI systems” AI or other digital technologies”. The UNECE has historically
and “objectives and approaches for explainability and developed and adopted standards under the “Working
interpretability of ML models and AI systems”. Party on Regulatory Cooperation and Standardization
Policies” (WP.6) (UNECE, 2024). In the context of the new
project, UNECE issued in November 2023 a document
B. ITU proposing various recommendations and approaches
on the regulation of AI-embedded products that related
The ITU’s Telecommunications Standardisation Sector to international trade in general, and WTO disciplines in
(ITU-T) has developed various technical standards on AI in particular, including that:
the form of frameworks for evaluating intelligence levels
of future networks and for data handling, as well as “Governments should ensure that regulatory measures
architectural frameworks for machine learning and AI-based applied to products with embedded digital technologies
networks.35 The ITU-T is one of the ITU branches that are consistent with the World Trade Organization (WTO)
develops international standards in the area of information Technical Barriers to Trade (TBT) Agreement. This includes,
and communication technologies.36 but is not limited to, the TBT Agreement’s obligations
pertaining to notification, publication, non-discrimination,
avoidance of unnecessary barriers to trade, achievement
C. Institute of Electrical and of legitimate objectives and use of international standards”
Electronics Engineers (IEEE) (UNECE, 2023).

The IEEE has developed various standards dealing with This project is still ongoing, and no outcome has yet been
socio-technical issues related to AI systems. Among other adopted with respect to the proposals in the UNECE
functions, the IEEE develops international standards on November 2023 document.

Endnotes
1 See Chapter 2(a) for the OECD’s definition of “AI system”. 5 It is important to note, however, that G20 economies did not
adhere to the definitional part of the OECD AI Principles (2019),
2 According to OECD AI Principles (2019a), section 1.I, an “AI
including the definitions of “AI system” and “AI system lifecycle”,
system lifecycle” involves: “i) ‘design, data and models’; which is
although they did not expressly reject this part either.
a context dependent sequence encompassing planning and
design, data collection and processing, as well as model building; 6 See [Link]
ii) ‘verification and validation’; iii) ‘deployment’; and iv) ‘operation [Link]#:~:text=We%2C%20the%20G20%20
and monitoring’. These phases often take place in an iterative Leaders%2C%20meeting,century%20for%20all%20by%20
manner and are not necessarily sequential. The decision to retire empowering; [Link]
an AI system from operation may occur at any point during the ROMELEADERSDECLARATION_0.pdf, and [Link]
operation and monitoring phase”. [Link]/Images/CPV/[Link].
3 See [Link] 7 See [Link]
4 See [Link]
osaka19/pdf/documents/en/annex_08.pdf.

102
 ANNEX

8 See [Link] 26 UN General Assembly (UNGA) Resolution “Seizing

Annex1_DECLARATION-OF-G20-DIGITAL-MINISTERS-2021_ the Opportunities of Safe, Secure and Trustworthy Artificial
[Link]. Intelligence Systems for Sustainable Development” (A/78/L.49),
adopted on 11 March 2024 (available at [Link]
9 See [Link]
[Link]/doc/undoc/ltd/n24/065/92/pdf/[Link]).
10 Unlike OECD (2021), the Recommendation does not define
27 The UN AI Resolution clarifies that its content is limited to
“AI”. It states that it “does not have the ambition to provide one AI systems in the “non-military domain”. It also clarifies that, for
single definition of AI, since such a definition would need to its purpose, the AI “lifecycle” comprises the following stages:
change over time, in accordance with technological developments. “pre-design, design, development, evaluation, testing, deployment,
Rather, its ambition is to address those features of AI systems use, sale, procurement, operation and decommissioning”
that are of central ethical relevance”. Nevertheless, it does ([Link]
provide a broad understanding of what “AI systems” mean, [Link]).
i.e. “systems which have the capacity to process data and
information in a way that resembles intelligent behaviour, and 28 See [Link]
typically includes aspects of reasoning, learning, perception, [Link].
prediction, planning or control”. 29 See [Link]
11 See [Link] and_ip.htm.
html and [Link] 30 See [Link]
document01_en.pdf.
31 ISO and IEC established a joint committee (JTC 1) to coordinate
12 See [Link] development of digital technology standards in 1987. See JTC 1
html. (2024).
13 See [Link] 32 See [Link]
u/0/w/0/d/0.
14 See [Link]
html. 33 See [Link]
u/1/w/0/d/0.
15 See [Link]
34 See [Link]
16 See [Link] iec:tr:24368:ed-1:v1:en.
2024/07/oecd-launches-pilot-to-monitor-application-of-
[Link]. 35 See ITU-T Y.3172 “Architectural framework for machine
learning in future networks including IMT -2020” (IMT stands
17 See [Link] for “International Mobile Telecommunications”) (ITU, 2020b);
digital-joint-ministerial-statement-15_10_24.pdf. ITU-T Y.3173 “Framework for evaluating intelligence levels of
18 See [Link] future networks including IMT-2020” (ITU, 2020d); ITU-T Y.3174
“Framework for data handling to enable machine learning in
page/field_ref_resources/390de76d-d4f5-4f45-a7b4-
future networks including IMT-2020” (ITU, 2020c); ITU-T Y.3177
f6879c30c389/0fbffe8a/20231201_en_news_g7_result_00.
“Architectural framework for artificial intelligence-based network
pdf, [Link]
automation for resource and fault management in future networks
[Link], [Link]
including IMT-2020” (ITU, 2020a).
[Link]/policy/economy/g20_summit/osaka19/en/documents/
final_g20_osaka_leaders_declaration.html and [Link] 36 See [Link]
[Link]/en/opc-news/speeches/2022/communique-g7-220908/.
37 See [Link]
19 See [Link] linkslist_text&utm_medium=lp-about&utm_campaign=at-a-glance.
ministers-digital-trade-principles.
20 See [Link]
Publikation/ E N /Others/G7_2023_Communique.pdf?__
blob=publicationFile&v=2 and [Link]
contents/node/basic_page/field_ref_resources/390de76d-d4f5-
4f45-a7b4-f6879c30c389/0fbffe8a/20231201_en_news_g7_
result_00.pdf.
21 See [Link] and
[Link]
[Link].
22 See [Link]
page/field_ref_resources/390de76d-d4f5-4f45-a7b4-
f6879c30c389/0fbffe8a/20231201_en_news_g7_result_
[Link].
23 See [Link]
24 “For the AI Safety Summit at Bletchley Park, frontier AIs were
defined as models that can perform a wide variety of tasks and
match or exceed the capabilities present in today’s most advanced
models” – Glossary, Interim Report (page 88).
25 See [Link]
66f5311f080bdf716392e922/international_scientific_report_
on_the_safety_of_advanced_ai_interim_report.pdf.

103
 ANNEX

Responses were received from Susan Aaronson (George

Annex 4 Survey of Washington University; Centre for International Governance
Innovation), Dan Ciuriak (Centre for International Governance
Academics Innovation), Johannes Fritz (Digital Policy Alert), Olia
Kanevskaia (Utrecht University, Department of International
and European Law, and Utrecht Centre for Regulation and
Following a literature review, scholars working at the Enforcement in Europe), Kholofelo Kugler (University of
intersection of trade and AI were identified and invited to Lucerne; Counsel, Advisory Centre on WTO Law (ACWL)),
respond to the following two questions. Heidi Lund (National Board of Trade Sweden), Petros
1. What are the main challenges that AI poses to the WTO’s Mavroidis (Columbia Law School), Hildegunn Kyvik Nordås
current rules, principles and practices? (Council on Economic Policies (CEP), Örebro University),
2. How could WTO members ensure such rules, principles Eduardo Paranhos (Head of AI Work Group at Associação
and practices remain fit-for-purpose in light of challenges? Brasileira das Empresas de Software (ABES)) and Shin-Yi
In your view, may specific rules, principles and practices Peng (National Tsing Hua University). Chapter 4(f) reflects
need to be adjusted, and if so how? these responses.

104
 BIBLIOGRAPHY

Bibliography
Aaronson, S.A. (2024a), "Data Disquiet: Concerns about the & Society 20(3):973–989. Available at [Link]
Governance of Data for Generative AI", SSRN Electronic 10.1177/1461444816676645
Journal. Available at [Link]
Association of Southeast Asian Nations (ASEAN) (2024),
Aaronson, S.A. (2024b), "The Age of AI Nationalism and its ASEAN Guide on AI Governance and Ethics, ASEAN.
Effects", SSRN Electronic Journal. Available at [Link] Available at [Link]
org/10.2139/ssrn.4803311 02/AS EAN-Guide-on-AI-Governance-and-Ethics_
beautified_201223_v2.pdf
Aaronson, S. A. (2019), "Data is different, and that’s why
the world needs a new approach to governing cross-bor- Assad, S., Clark, R., Ershov, D. and Xu, L. (2024),
der data flows", Digital Policy, Regulation and Governance "Algorithmic Pricing and Competition: Empirical Evidence
21(5):441–460. Available at [Link] from the German Retail Gasoline Market", Journal of
DPRG-03-2019-0021 Political Economy 132(3):723–771. Available at https://
[Link]/10.1086/726906
Aaronson, S. A. (2023), "Building trust in digital trade will
require a rethink of trade policy-making", Oxford Review of Auda-Nepad (2023), "AUDA-NEPAD White Paper:
Economic Policy 39(1):98–109. Available at: [Link] Regulation and Responsible Adoption of AI in Africa Towards
org/10.1093/oxrep/grac046 Achievement of AU Agenda 2063", DigWatch. Available
at [Link]
Aaronson, S. A. and Leblond, P. (2018), "Another Digital regulation-and-responsible-adoption-of-ai-in-africa-
Divide: The Rise of Data Realms and its Implications for towards-achievement-of-au-agenda-2063
the WTO", Journal of International Economic Law 21(2):
245–272. Available at [Link] Autor, D. (2022), "The Labor Market Impacts of Technological
Change: From Unbridled Enthusiasm to Qualified Optimism
Acemoglu, D. (2024), "The Simple Macroeconomics of AI", to Vast Uncertainty", NBER Working Paper No. 30074,
NBER Working Paper No. 32487 Cambridge (MA): National Cambridge (MA): National Bureau of Economic Research
Bureau of Economic Research (NBER). Available at https:// (NBER). Available at [Link]
[Link]/sites/default/files/2024-04/The%20
Simple%20Macroeconomics%20of%[Link] Autor, D. (2024), "Applying AI to Rebuild Middle Class Jobs",
NBER Working Paper No. 32140, Cambridge (MA): Na-
Agrawal, A., Gans, J. and Goldfarb, A. (2019), The econom- tional Bureau of Economic Research (NBER). Available at
ics of artificial intelligence: An agenda, Chicago: The Univer- [Link]
sity of Chicago Press.
Baldwin, R. (2024), "AI's new skills twist", IMD. Available at
Alania, A., Firdaus, A., Callegari, A., Hana, A., Shell, S. and [Link]
Ohkura, Y. (2022), "Looking Ahead: The Role of Standards in skills-twist/
the Future of Artificial Intelligence (AI) Governance", London:
University College London. Barteková, E. and Börkey, P. (2022), "GHG Emission Trends
and Targets (GETT): Harmonised quantification methodolo-
Alonso, C., Berg, A., Kothari, S., Papageorgiou, C. and Re- gy and indicators", OECD Environment Working Paper No.
hman, S. (2022), "Will the AI revolution cause a great diver- 230, Paris: Organisation for Economic Co-operation and
gence?", Journal of Monetary Economics 127:18–37. Avail- Development (OECD). Available at [Link]
able at [Link] decef216-en

Altana (2021), "How Value Chain Visibility Protects Crucial Begenau, J., Farboodi, M. and Veldkamp, L. (2018), "Big data
Vaccine Supplies", Altana. Available at [Link] in finance and the growth of large firms", Journal of Monetary
[Link]/altana/216c2ef0-51d5-420e-a413-37e6aa991edf_ Economics 97:71–87. Available at [Link]
How+Value+Chain+Visibility+Protects+Crucial+Vaccine+- jmoneco.2018.05.013
[Link]
Bello y Villarino, J. M. (2023), "Money, Power and AI" in
Amodei, D. and Hernandez, D. (2024), "AI and compute", Bednarz, Z. and Zalnieriute, M. (eds. 1st). Cambridge
OpenAI. Available at [Link] (UK): Cambridge University Press. Available at [Link]
and-compute org/10.1017/9781009334297

Ananny, M. and Crawford, K. (2018), "Seeing with- Bonawitz, K., Eichner, H., Grieskamp, W., Huba, D.,
out knowing: Limitations of the transparency ideal and Ingerman, A., Ivanov, V., Kiddon, C., Kone ný, J., Maz-
its application to algorithmic accountability", New Media zocchi, S., McMahan, H. B., Van Overveldt, T., Petrou,

105
 BIBLIOGRAPHY

D., Ramage, D. and Roselander, J. (2019), "Towards Ottawa: Edward Elgar. Available at [Link]
Federated Learning at Scale: System Design", Cornell com/shop/gbp/research-handbook-on-digital-trade-
University, arXiv:1902.01046. Available at [Link] [Link]
abs/1902.01046
Ciuriak, D. (2024), "Technological Conditions and the Rise
Borchert, I., Gootiiz, B., Grover Goswami, A. and Mat- and Fall of the Rules-Based System", SSRN Electronic
too, A. (2017), "Services Trade Protection and Economic Journal. Available at [Link]
Isolation", The World Economy 40(3):632–652. Available at
[Link] Council of Europe (2024), Framework Convention on
Artificial Intelligence, Human Rights, Democracy and the
BP (2022), "BP Statistical Review of World Energy", Rule of Law, France: Council of Europe. Available at https://
London (UK): BP [Link]/ai-convention-brochure/1680afaeba

Bradford, A. (2023), "Whose AI Revolution?" Project Csernatoni, R. (2024), "Charting the Geopolitics and
Syndicate. Available at [Link] European Governance of Artificial Intelligence", Carneg-
onpoint/ai-regulation-us-eu-china-challenges- ie Endowment for International Peace. Available at https://
opportunities-by-anu-bradford-2023-09 [Link]/files/Csernatoni_-_Governance_
[Link]
Brynjolfsson, E., Hui, X. and Liu, M. (2019), "Does Machine
Translation Affect International Trade? Evidence from a Degot, C., Duranton, S., Fredeau, M. and Hutchinson, R.
Large Digital Platform", Management Science 65(12):5449– (2021), "Reduce Carbon Costs with the Power of AI",
5460. Available at [Link] Boston Consultancy Group. Available at [Link]
com/publications/2021/ai-to-reduce-carbon-emissions
Brynjolfsson, E., Li, D. and Raymond, L. R. (2023), "Generative
AI at Work" NBER Working Paper No. 31161 Cambridge Del Giovane, C., López González, J. and Ferencz, J. (2023),
(MA): National Bureau of Economic Research (NBER). "The Nature, Evolution and Potential Implications of Data
Available at [Link] Localisation Measures", OECD Trade Policy Paper No.
278, Paris: Organisation for Economic Co-operation and
Cambridge University (2023), "Voluntary technical standards Development (OECD). Available at [Link]
for AI and robotics", Global AI Governance. Available at 179f718a-en
[Link]
Dell’Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf,
Castelvecchi, D. (2016), "Can we open the black box of H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. and
AI?", Nature 538(7623):20–23. Available at [Link] Lakhani, K. R. (2023), "Navigating the Jagged Technolog-
org/10.1038/538020a ical Frontier: Field Experimental Evidence of the Effects of
AI on Knowledge Worker Productivity and Quality", SSRN
Cazzaniga, M., Jaumotte, F., Li, L., Melina, G., Panton, A., Electronic Journal. Available at [Link]
Pizzinelli, C., Rockall, E. J. and Tavares, M. M. (2024), ssrn.4573321
"Gen-AI: Artificial Intelligence and the Future of Work",
International Monetary Fund. Available at [Link] Dentons, Spatial Web Foundation and Verses (2023),
[Link]/en/Publications/Staff-Discussion-Notes/Issues/ "The Future of Global AI Governance", Dentons. Available
2024/01/14/Gen-AI-Artificial-Intelligence-and-the- at [Link]
Future-of-Work-542379 27/-/media/[Link]

Cernat, L. (2023), "The Art of the Mini-Deals: The Invisible Ebers, M. (2024), "Truly Risk-Based Regulation of Artificial
Part of EU Trade Policy", Policy Brief No.11/2023, Brussels: Intelligence—How to Implement the EU’s AI Act", SSRN
European Centre for International Political Economy (ECIPE). Electronic Journal. Available at [Link]
ssrn.4870387
Cesario, A., D’Oria, M., Auffray, C. and Scambia, G. (Eds.)
(2023), Personalized medicine meets artificial intelligence: Ebers, M. and Navas Navarro, S. (2020), Algorithms and law,
Beyond “hype”, towards the metaverse, Switzerland: Cambridge: Cambridge University Press.
Springer.
Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2023),
Chen, Z. (2022), "The AI revolution is underway, and this "GPTs are GPTs: An Early Look at the Labor Market Impact
is good news for Customs", WCO News. Available at https:// Potential of Large Language Models", Cornell University
[Link]/magazine/wco-news-99-issue-3-2022/ arXiv:2303.10130. Available at [Link]
ai-revolution-is-underway/ 10130

Ciuriak, D. (2023), "Digital Economy Agreements: Where Epoch AI (2024), "Affiliation of research teams building
Do We Stand and Where Are We Going?", in Collins, D. notable AI systems, by year of publication", Epoch AI.
and Geist, M. (eds.). Research Handbook on Digital Trade, Available at [Link]

106
 BIBLIOGRAPHY

European Commission (2022), The Digital Markets Act: National Bureau of Economic Research (NBER). Available at
Ensuring fair and open digital markets, Brussels: Europe- [Link]
an Commission. Available at [Link]
eu/strategy-and-policy/priorities-2019-2024/europe- Goldfarb, A. and Tucker, C. (2012), "Shifts in Privacy
fit-digital-age/digital-markets-act-ensuring-fair-and- Concerns", American Economic Review 102(3):349–353.
open-digital-markets_en Available at [Link]

European Commission (2024b), European Data Governance Goldman Sachs (2023), The Potentially Large Effects of
Act, Brussels: European Commission. Available at https:// Artificial Intelligence on Economic Growth, New York:
d i g i t a l - s t r a t e g y. e c . e u r o p a . e u / e n / p o l i c i e s / d a t a - Goldman Sachs. Available at [Link]
governance-act com/content/research/en/reports/2023/03/27/d64e052b-
[Link]
European Commission, CEPS, ICF and Wavestone (2021),
"Study to support an impact assessment of regulatory Gornet, M. and Maxwell, W. (2024), "The European
requirements for Artificial Intelligence in Europe", approach to regulating AI through technical standards",
European Commission Final Report No. D5, Brussels: Internet Policy Review 13(3). Available at [Link]
European Commission (EC) Publications Office. Available at 10.14763/2024.3.1784
[Link]
Goyens, M. (2020), "Effective Consumer Protection
European Parliament (2024), Artificial Intelligence Act, Frameworks in a Global and Digital World", Journal of
Strasbourg: European Parliament. Available at [Link] Consumer Policy 43(1):195–207. Available at [Link]
[Link]/doceo/document/TA-9-2024-0138_ org/10.1007/s10603-019-09423-2
[Link]
Head, K & Ries, J. (2001), "Increasing Returns versus
European Union (2024), EU AI Act, European Union. National Product Differentiation as an Explanation for the
Available at [Link] Pattern of U.S.-Canada Trade", American Economic Review,
PE-24-2024-INIT/en/pdf 91 (4): 858–876.

Financial Times (2022), "China's chip industry set for deep Hill, P. (1999), "Tangibles, Intangibles and Services: A New
pain from US export controls", Financial Times. Available at Taxonomy for the Classification of Output", The Canadian
[Link] Journal of Economics / Revue Canadienne d’Economique
ece71d2cb267 32(2):426. Available at [Link]

Firooz, H., Liu, Z. and Wang, Y. (2022), "Automation, Market Holmes, C. (2024), "How upcoming legislation will harness
Concentration, and the Labor Share", FRBSF Working algorithms, AI, and big tech for public good", Finextra.
Paper No. 2022-05, San Francisco: Federal Reserve Bank Available at [Link]
of San Francisco (FRBSF). Available at [Link] how-upcoming-legislation-will-harness-algorithms-ai-and-
org/economic-research/publications/working- big-tech-for-public-good
papers/2022/05/
Holzer, K. (2019), "Addressing Tensions and Avoiding
Fritz, J. and Giardini, T. (2024), "Emerging Contours of AI Disputes: Specific Trade Concerns in the TBT Commit-
Governance and the Three Layers of Regulatory Heterogene- tee", Global Trade and Customs Journal 14(3):102–116.
ity", Digital Policy Alert Working Paper No. 24-001, St. Gallen: Available at [Link]
Digital Policy Alert. Available at: [Link]
[Link]/reports/1715797071525_ Horn, H., Mavroidis, P. C. and Wijkström, E. N. (2013), "In the
Fritz%20&%20Giardini%20-%20AI%20Governance% Shadow of the DSU: Addressing Specific Trade Concerns in
20&%203%20Layers%20of%20Regulatory%20 the WTO SPS and TBT Committees", Journal of World Trade
[Link] 47(4):729–759. Available at [Link]
TRAD2013024
Gal, M. and Elkin-Koren, N. (2017), "Algorithmic Consumers",
Harvard Journal of Law and Technology 30. Available at Hoyer Gosselink, B., Brandt, K., Croak, M., DeSalvo, K.,
[Link] Gomes, B., Ibrahim, L., Johnson, M., Matias, Y., Porat, R.,
Walker, K. and Manyika, J. (2024), "AI in Action: Accelerat-
Gmyrek, P., Berg, J. and Bescond, D., (2023). Generative ing Progress Towards the Sustainable Development Goals",
AI and jobs: A global analysis of potential effects on job Google Research Brief, San Francisco: Google. Available
quantity and quality. ILO Working Paper, 96. Available at: at [Link]
[Link] google/en//resources/[Link]
global-analysis-potential-effects-job-quantity-and
Hunt, K. (2020), "Safer Connected Consumer Products",
Goldfarb, A. and Trefler, D. (2018), "AI and International ASTM International. Available at [Link]
Trade", NBER Working Paper No. 24254 Cambridge (MA): [Link]

107
 BIBLIOGRAPHY

International Electrotechnical Commission (IEC) (2023), (UNESCO) (2013), The State of Broadband 2013:
What are the latest tech trends?, Geneva: IEC. Available at Universalizing Broadband, Geneva: ITU and UNESCO.
[Link] Available at [Link]
[Link]=en
Institute of Electrical and Electronics Engineer (IEEE)
(2023), Algorithmic Bias Considerations, New Jersey: IEEE. Ishai, G., -B., Dean, J., Manyika, J., Porat, R., Varian, H.,
Available at [Link] and Walker, K. (2024), "AI and the Opportunity for Shared
Prosperity: Lessons from the History of Technology and the
International Monetary Fund (IMF), Organisation for Economy", Cornell University arXiv:2401.09718. Available at
Economic Co-operation and Development (OECD), [Link]
World Bank and World Trade Organization (WTO) (2023),
Digital Trade for Development, Washington, D.C., Paris Jones, E. (2023), "Digital disruption: Artificial intelligence
and Geneva: IMF, OECD, World Bank and WTO. and international trade policy", Oxford Review of Economic
Available at [Link] Policy 39(1):70–84. Available at [Link]
dtd2023_e.pdf oxrep/grac049

Intellectual Property Repository (2023), Judgment of the Jones, E., Kira, B. and Tavengerwei, R. (2024), "Norm
first case of copyright infringement of AI-generated images, Entrepreneurship in Digital Trade: The Singapore-led
Intellectual Property Rights. Available at [Link] Wave of Digital Trade Agreements", World Trade Review
[Link]/s/Wu3-GuFvMJvJKJobqqq7vQ 23(2):208–241. Available at [Link]
S1474745624000089
International Business Machines Corporation (2024), "IBM
Global AI Adoption Index 2023", IBM. Available at https:// International Standards Organization (ISO) and International
[Link]/players/English/9240059-ibm-2023- Electrotechnical Commission (IEC) Joint Technical
global-ai-adoption-index-report/ Committee 1 (2024), ISO and IEC Joint Technical
Committee, Geneva: ISO and IEC. Available at https://
International Energy Agency (2024), "Electricity 2024", [Link]/
Paris: International Energy Alliance. Available at [Link]
[Link]/reports/electricity-2024 Judge, B., Nitzberg, M. and Russell, S. (2024), "When code
isn’t law: Rethinking regulation for artificial intelligence", Policy
International Telecommunications Union (ITU) (2020a), and Society, puae020. Available at [Link]
Architectural framework for artificial intelligence-based polsoc/puae020
network automation for resource and fault management
in future networks including IMT-2020, Geneva: ITU. Karttunen, M. B. (2020), Transparency in the WTO SPS and
Available at [Link] TBT Agreements: The real jewel in the crown, Cambridge:
[Link]?rec=14598 Cambridge University Press.

International Telecommunications Union (ITU) (2020b), Archi- Kerry, C. F. (2024), "Small Yards, Big Tents: How to
tectural framework for machine learning in future networks in- Build Cooperation on Critical International Standards",
cluding IMT-2020, Geneva: ITU. Available at [Link] Brookings Institute (Governance Studies), Washington DC:
int/ITU-T/recommendations/[Link]?rec=13894&lang=en Brookings Institute. Available at [Link]
wp-content/uploads/2024/03/GS_03062024_
International Telecommunications Union (ITU) (2020c), [Link]
Framework for data handling to enable machine learning in
future networks including IMT-2020, Geneva: ITU. Available Kilic, B. (2024), "Into Uncharted Waters: Trade Secrets
at [Link] Law in the AI Era" Centre for International Governance
14134&lang=en Innovation (CIGI) Paper No- 295, Waterloo: CIGI. Available
at [Link]
International Telecommunications Union (ITU) (2020d),
Framework for evaluating intelligence levels of future net- Klein, E. and Stewart, P. (2024), "Envisioning a Global
works including IMT-2020, Geneva: ITU. Available at https:// Regime Complex to Govern Artificial Intelligence", Carne-
[Link]/ITU-T/recommendations/[Link]?rec=14133 gie Endowment for International Peace, Washington DC:
Carnegie Endowment for International Peace. Available
International Telecommunications Union (ITU) (2023), Glob- at [Link]
al Offline population steadily declines to 2.6 billion people Regime_Complex.pdf
in 2023, Geneva: ITU. Available at [Link]
reports/statistics/2023/10/10/ff23-internet-use/acts and Krummenacher, P. (2023), "International Trade and
Figures 2023 - Internet use ([Link]) Artificial Intelligence: Is trade policy ready for Chat GPT?"
International Institute for Sustainable Development (IISD).
International Telecommunications Union (ITU) and United Available at [Link]
Nations Educational, Scientific and Cultural Organization international-trade-artificial-intelligence-chatgpt

108
 BIBLIOGRAPHY

Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Maskus, K. and Reichman, J. (2004), "The Globalization
Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Of Private Knowledge Goods And The Privatization Of
Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Global Public Goods", Journal of International Economic Law
Stott, J., Pritzel, A., Mohamed, S. and Battaglia, P. (2023), 7(2):279–320.
"GraphCast: Learning skillful medium-range global weather
forecasting", Cornell University arXiv:2212. 12794. Available Maslej, N., Fattorini, L., Perrault, R., Parli, V., Reuel, A., Bryn-
at [Link] jolfsson, E., Etchemendy, J., Ligett, K., Lyons, T., Manyika, J.,
Niebles, J. C., Shoham, Y., Wald, R. and Clark, J. (2024), "The
Lee, S. (2024), "The Impact of Artificial Intelligence on Small AI Index 2024 Annual Report", Institute for Human-Centered
Businesses", Harvard Model Congress Boston. Available AI, Stanford: Stanford University.
at [Link]
0c9145fa68863e/t/6549741749fcc212c2826d Mavroidis, P. (2016), The Regulation of International Trade,
2c/1699312663764/HMC2024_Senate_SBE_1.pdf Cambridge (MA): MIT Press.

Lee-Makiyama, H. (2018), "Briefing note: AI & Trade Policy", Mayer-Schönberger, V. and Ramge, T. (2018), Reinventing
Tallinn Digital Summit, Tallinn: Tallinn Digital Summit. capitalism in the age of big data (eds.), London: Basic Books.
Available at [Link]
TDS2018-BriefingNote_AI_Trade_Policy.pdf McDaniels, D., Molina, A. C. and Wijkström, E. (2018), "How
Does the Regular Work of WTO Influence Regional Trade
Lim, A. H. (2021), "Trade Rules for Industry 4.0: Why the Tech- Agreements?" WTO Working Paper No. 2018/06, Geneva:
nical Barriers to Trade Agreement Matters Even More", in Peng, World Trade Organization (WTO). Available at [Link]
S., Lin, C-F., Streinz, T., Artificial Intelligence and Internation- org/10.30875/10cbd249-en
al Economic Law, Cambridge: Cambridge University Press.
Available at [Link] McKinsey (2022), "Localization of data privacy regulations
creates competitive opportunities", Mckinsey. Available at
Lin, C.-F. (2021), "Public Morals, Trade Secrets, and the [Link]
Dilemma of Regulating Automated Driving Systems" in Peng, our-insights/localization-of-data-privacy-regulations-
S., Lin, C-F., Streinz, T., Artificial Intelligence and International creates-competitive-opportunities
Economic Law, Cambridge: Cambridge University Press.
Available at [Link] McKinsey (2023), "Economic potential of generative
identifier/9781108954006%23CN-bp-5/type/book_part AI", McKinsey. Available at [Link]
capabilities/mckinsey-digit al/our-insights/the-
Liu, H. W. and Lin, C. F. (2020), "Artificial Intelligence and economic-potential-of-generative-AI-the-next-productivity-
Global Trade Governance: A Pluralist Agenda", Harvard frontier#introduction
International Law Journal 61(2).
Melitz, J. and Toubal, F. (2014), “Native language, spoken
López-González, J., Sorescu, S. and Kaynak, P. (2023), "Of language, translation and trade”, Journal of International
bytes and trade: Quantifying the impact of digitalisation on Economics 93(2).
trade", OECD Trade Policy Paper No. TAD/TC/WP(2022)11/
FINAL, Paris: Organisation for Economic Co-operation and Meltzer, J. P. (2023), "Toward international cooperation on
Development (OECD). Available at [Link] foundational AI models", Global Economy and Development-
document/TAD/TC/WP(2022)11/FINAL/en/pdf Brookings Institute. Available at [Link]
e d u / w p - c o n t e n t / u p l o a d s / 2 0 2 3 / 1 1 / Fo u n d a t i o n a l -
Luintel, K. B. and Khan, M. (2009), "Heterogeneous Ideas AI-Models_Meltzer.pdf
Production and Endogenous Growth: An Empirical
Investigation", The Canadian Journal of Economics 42(3): Miller, C. (2022), Chip war: The fight for the world’s most
1176–1205. critical technology (eds.), New York: Scribner.

Lund, B. D., Wang, T., Mannuru, N. R., Nie, B., Shimray, Mishra, N. (2024), International trade law and global data
S. and Wang, Z. (2023), "ChatGPT and a New Academic governance: Aligning perspectives and practices, Oxford:
Reality: Artificial Intelligence-Written Research Papers Hart Publishing.
and the Ethics of the Large Language Models in Scholarly
Publishing", Journal of the Association for Information Mitchell, A. D., Let, D. and Tang, L. (2023), "AI Regulation
Science and Technology 74(5):570–581. Available at and the Protection of Source Code", International Journal
[Link] of Law and Information Technology 31(4):283–301.
Available at [Link]
Manyika, J., Lund, S., Bughin, J., Woetzel, J., Stamenov, K.
and Dhringra, D. (2016), "Digital Globalization: The New Mitchell, A. D. and Mishra, N. (2018), "Data at the Docks:
Era of Global Flows", McKinsey Global Institute. Available Modernising International Trade Law for the Digital
at [Link] Economy", Vanderbilt Journal of Entertainment & Technology
2016/06/709- mgi_digital_globalization.pdf Law 20.

109
 BIBLIOGRAPHY

Moon, S. (2008), "Does TRIPS Art. 66.2 Encourage OECD. Available at [Link]
Technology Transfer to LDCs? An Analysis of Country instruments/OECD-LEGAL-0449
Submissions to the TRIPS Council (1999-2007)",
UNCTAD-ICTSD Project on IPRs and Sustainable Organisation for Economic Co-operation and Development
Development Policy Brief No. 2, Geneva: United Nations (OECD) (2019b), Open Government Data, Paris: OECD.
Conference on Trade and Development. Available at Available at [Link]
[Link] numerique/[Link]
pb20092_en.pdf
Organisation for Economic Co-operation and Development
Morley, J., Floridi, L., Kinsey, L. and Elhalal, A. (2020), "From (OECD) (2021a), Competition and AI. In OECD Business
What to How: An Initial Review of Publicly Available AI and Finance Outlook 2021: AI in Business and Finance,
Ethics Tools, Methods and Research to Translate Paris: OECD. Available at [Link]
Principles into Practices", Science and Engineering Ethics ba682899-en
26(4):2141–2168. Available at [Link]
s11948-019-00165-5 Organisation for Economic Co-operation and Development
(OECD) (2021b), "Data portability, interoperability and
Morris, M. R., Sohl-dickstein, J., Fiedel, N., Warkentin, T., digital platform competition", OECD Competition
Dafoe, A., Faust, A., Farabet, C. and Legg, S. (2024), Committee Discussion Paper No. DAF/COMP/WD(2021)44,
"Levels of AGI: Operationalizing Progress on the Path to Paris: OECD.
AGI", Cornell University arXiv:2311.02462 arXiv. Available
at [Link] Organisation for Economic Co-operation and Development
(OECD) (2021c), Putting the OECD AI Principles into
Muralidharan, K., Singh, A. and Ganimian, A.J. (2019), practice: Progress and future perspectives, Paris: OECD.
“Disrupting Education? Experimental Evidence on Available at [Link]
Technology-Aided Instruction in India”, American Economic
Review, 109(4): 1426–1460. Organisation for Economic Co-operation and Development
(OECD) (2022a), "Artificial Intelligence and international
National Board of Trade Sweden (2023), "Innovation, trade: Some preliminary implications", OECD Trade Policy
AI, Technical Regulation and Trade". Available at https:// Paper No. 260, Paris: OECD. Available at [Link]
[Link]/globalassets/publikationer/ org/10.1787/13212d3e-en
rapporter/2023/innovation-ai-technical-regulation-and-
[Link] Organisation for Economic Co-operation and Development
(OECD) (2022b), "Measuring the environmental impacts
Nicoletti, G., Von Rueden, C. and Andrews, D. (2020), of artificial intelligence compute and applications", OECD
"Digital technology diffusion: A matter of capabili- Digital Economy Paper No. 361, Paris: OECD. Available at
ties, incentives or both?", European Economic Review [Link]
128:103513. Available at [Link]
[Link].2020.103513 Organisation for Economic Co-operation and Development
(OECD) (2023a), Measuring the Internet of Things,
Nordås, H. K. and Rouzet, D. (2017), "The Impact of Services Paris: OECD. Available at [Link]
Trade Restrictiveness on Trade Flows", The World 021333b7-en
Economy 40(6):1155–1183. Available at [Link]
org/10.1111/twec.12424 Organisation for Economic Co-operation and Development
(OECD) (2023b), "Raw materials critical for the green
Noy, S. and Zhang, W. (2023), "Experimental evidence on transition: Production, international trade and export
the productivity effects of generative artificial intelligence", restrictions", OECD Trade Policy Paper No. 269, Paris:
Science 381(6654):187–192. Available at [Link] OECD. Available at [Link]
org/10.1126/science.adh2586
Organisation for Economic Co-operation and Develop-
Organisation for Economic Co-operation and Development ment (OECD) (2023c), Regulatory Sandboxes in Artificial
(OECD) (2016a), Economic and social benefits of internet Intelligence, Paris: OECD. Available at [Link]
openness, Paris: OECD. Available at [Link] 10.1787/8f80a0e6-en
10.1787/5jlwqf2r97g5-en
Organisation for Economic Co-operation and Development
Organisation for Economic Co-operation and Development (OECD) (2024a), "Explanatory memorandum on the
(OECD) (2016b), OECD Science, Technology and updated OECD definition of an AI system", OECD Artificial
Innovation Outlook 2016, Paris: OECD. Available at OECD. Intelligence Paper No. 8, Paris: OECD. Available at https://
[Link] [Link]/10.1787/623da898-en

Organisation for Economic Co-operation and Development Organisation for Economic Co-operation and Develop-
(OECD) (2019a), OECD AI Principles overview, Paris: ment (OECD) (2024b), "Using AI in the workplace", OECD

110
 BIBLIOGRAPHY

Artificial Intelligence Paper No. 11, Paris: OECD. Available at public deserves to know", [Link] Policy Observatory.
[Link] Available at [Link]
ai-consume
Organisation for Economic Co-operation and Development
(OECD) and World Trade Organization (WTO) (2024), A Rozite, V., Miller, J., and Miller, S. (2023), "Why AI and energy
Song of Data Flows and Trust – A Quantitative Analysis of are the new power couple", International Energy Agency.
Cross-border Data Flow Regulation, Paris: OECD. Available at [Link]
energy-are-the-new-power-couple
Organisation for Economic Co-operation and Development
(OECD) and World Trade Organization (WTO) (2019), Sastry, G., Heim, L., Belfield, H., Anderljung, M., Brundage,
"Facilitating Trade Through Regulatory Cooperation: M., Hazell, J., O’Keefe, C., Hadfield, G. K., Ngo, R., Pilz, K.,
The case of the WTO’s TBT and SPS Agreements and Gor, G., Bluemke, E., Shoker, S., Egan, J., Trager, R. F., Avin,
Committees", Geneva: WTO. Available at [Link] S., Weller, A., Bengio, Y. and Coyle, D. (2024), "Comput-
org/english/res_e/booksp_e/tbtsps19_e.pdf ing Power and the Governance of Artificial Intelligence
(arXiv:2402.08797)",Cornell University arXiv. Available at
Ossa, R. (2023), "Digital trade is key to boosting growth in [Link]
developing economies", World Trade Organization, Geneva:
WTO. Available at [Link] Seong, J., White, O., Woetzel, L., Smit, S., Devesa, T.,
ce_ralph_ossa_e/blog_ro_15dec23_e.htm Birshan, M. and Samandari, H. (2022), "Global flows:
The ties that bind in an interconnected world", McKinsey
Owoyemi, A., Owoyemi, J., Osiyemi, A. and Boyd, A. (2020), Institute Discussion Paper, New York: McKinsey.
"Artificial Intelligence for Healthcare in Africa", Frontiers in
Digital Health 2:6. Available at [Link] Shadbolt, N. (2022), “From So Simple a Beginning”:
fdgth.2020.00006 Species of Artificial Intelligence", Daedalus 151(2):28–42.

Parkin, B. and Kay, C. (2024), "AI could kill off most call Smuha, N. (2024), "The European Union’s AI Act: Beyond
centres, says Tata Consultancy Services head", Financial motherhood and apple pie?", in The Cambridge Handbook
Times. Available at [Link] on the Law, Ethics and Policy of Artificial Intelligence,
ea71-42b0-b85b-86073354fb73 Cambridge: Cambridge University Press.

Pasquale, F. (2015), The black box society: The secret Stanford University (2023), "The Stanford Emerging
algorithms that control money and information, Cambridge Technology Review 2023", Stanford University, Stanford:
(MA): Harvard University Press. Stanford University. Available at [Link]
edu/sites/default/files/2023-11/SETR_web_231120.pdf
Possada, K. C., Ganne, E. and Piermartini, R. (2022), "The
Role of WTO Committees through the Lens of Specific Streinz, T. (2021), "International Economic Law’s
Trade Concerns Raised in the TBT Committee", World Trade Regulation of Data as a Resource for the Artificial
Review 21(4):411–431. Available at [Link] Intelligence Economy", in Peng, S., Lin, C.-F., and Streinz, T.
S1474745621000616 (eds.), Artificial Intelligence and International Economic
Law, Cambridge: Cambridge University Press. Available at
Pouget, H. (2023), "What will the role of standards be in AI [Link]
governance?", Ada Lovelace Institute. Available at https://
[Link]/blog/role-of-standards-in- Suleyman, M. and Bhaskar, M. (2023), The coming wave:
ai-governance/#_ftnref14 Technology, Power, and the Twenty-First Century’s Greatest
Dilemma, New York: Crown.
Prüfer, J. (2020), "Competition Policy and Data Sharing on
Data-driven Markets", Freidrich-Ebert-Stiftung, Geneva: Sun, R. and Trefler, D. (2023), "The Impact of AI and
Freidrich-Ebert-Stiftung. Available at [Link] Cross-Border Data Regulation on International Trade
pdf-files/fes/[Link] in Digital Services: A Large Language Model", NBER
Working Paper No. 31925, Cambridge (MA): National
Quigley, B. (2024), "Improving connectivity and accelerating Bureau of Economic Research (NBER). Available at https://
growth across Africa with new investments", Google. [Link]/10.3386/w31925
Available at [Link]/blog/products/infrastructure/
investing-in-connectivity-and-growth-for-africa. S & S Insider (2024), “AI Chip Market Expected
to Reach USD 621.15 Billion by 2032, at Rising
Raul, A. C. and Mushka, A. (2024), "The U.S. Plans to ‘Lead CAGR of 29.4%”, GlobalNewswire, 28 October 2024.
the Way’ on Global AI Policy", Lawfare. Available at https:// Available at: [Link]
[Link]/article/the-u.s.-plans-to-lead-the- release/2024/10/28/2970177/0/en/AI-Chip-Market-
way-on-global-ai-policy Expected-to-Reach-U S D-621-15-Billion-by-2032-
a t - R i s i n g - CAG R - o f - 2 9 - 4 - E x c l u s i v e - R e p o r t - b y - S -
Ren, S. (2023), "How much water does AI consume? The [Link]

111
 BIBLIOGRAPHY

Svanberg, M., Li, W., Fleming, M., Goehring, B. and Methodology, Paris: UNESCO. Available at [Link]
Thompson, N. (2024), Beyond AI Exposure: Which Tasks [Link]/ethics-ai/en/ram
are Cost-Effective to Automate with Computer Vision?,
MIT FutureTech Working Paper, Cambridge (MA): United Nations Educational, Scientific and Cultural
FutureTech. Available at [Link] Organization (UNESCO) (2023b), UNESCO’s
[Link]/2024-01-18+Beyond_AI_ Recommendation on the Ethics of Artificial Intelligence:
[Link] Key facts, Paris: UNESCO. Available at [Link]
[Link]/ark:/48223/pf0000385082
The Economist (2024), "AI holds tantalising promise for the
emerging world", The Economist. United Nations Office for Disaster Risk Reduction (UNDRR)
(2023), "Existential Risk and Rapid Technological Change",
The White House (2024), "What Drives the U.S. Services United Nations Office for Disaster Risk reduction,
Trade Surplus? Growth in Digitally-Enabled Services Geneva: UNDRR. Available at [Link]
Exports", The White House. Available at [Link] [Link]/media/86500/download?st art
[Link]/briefing-room/presidential-actions/2023/10/30/ Download=20240425
executive-order-on-the-safe-secure-and-trustworthy-devel-
opment-and-use-of-artificial-intelligence/ Verhulst, S. and Young, A. (2017), Open data in developing
economies: Toward building an evidence base on what works
United Nations (UN) (2024), Final Report: Governing AI and how, Cape Town: African Minds Publishers.
for Humanity, New York: UN. Available at governing_ai_for_
humanity_final_report_en.pdf Webb, A. and Hessel, A. (2022), The genesis machine: Our
quest to rewrite life in the age of synthetic biology (eds.),
United Nations Conference on Trade and Development New York: Public Affairs.
(UNCTAD) (2021a), Data Protection and Privacy Legislation
Worldwide, Geneva: UNCTAD. Available at [Link] Wells, S. (2023), "The “Trolley Problem” Doesn’t Work for
org/page/data-protection-and-privacy-legislation-worldwide Self-Driving Cars", IEEE Spectrum. Available at https://
[Link]/av-trolley-problem
United Nations Conference on Trade and Development
(UNCTAD) (2021b), Digital Economy Report 2021, Geneva: West, S. M. (2023), "Competition authorities need to move
UNCTAD. Available at [Link] fast and break up AI", Financial Times. Available at https://
economy-report-2021 [Link]/content/638b5be7-fab7-4fe6-a0cf-7dabef-
cdd722
United Nations Economic Commission for Europe (UNECE)
(2021), Artificial Intelligence Demystified, Geneva: UNECE. Wolff, A. (2022), "The Biden/Xi Meeting in Bali- What Was
Available at [Link] at Stake?", Pearson Institute for International Economics.
ECE_TRADE_C_CEFACT_2021_19E-AI_0.pdf Available at The Biden/Xi Meeting in Bali – What Was at
Stake? | PIIE
United Nations Economic Commission for Europe (UNECE)
(2022), White Paper on IoT Standards for Trade Facilitation, World Bank (2023), Trade Has Been a Powerful Driver of
Geneva: UNECE. Available at [Link] Economic Development and Poverty Reduction, Washington,
files/2023-10/WhitePaper_IoT_TradeFacilitation.pdf D.C.: World Bank.

United Nations Economic Commission for Europe (UNECE) World Bank (2023), Connecting to Compete 2023: Trade
(2023), "The regulatory compliance of products with Logistics in an Uncertain Global Economy – The Logistics
embedded artificial intelligence or other digital technologies", Performance Index and Its Indicators. Washington, D.C.:
UNECE Report No. ECE/CTCD/WP.6/2023/9, World Bank
Geneva: UNECE. Available at [Link]
default/files/2023-12/2023-9-Compliance-AI_Eng.pdf World Bank and World Trade Organization (WTO) (2023),
"Trade in services for development", WTO-WBG Paper,
United Nations Economic Commission for Europe (UNECE) Geneva: WTO. Available at [Link]
(2024), Working Party on Regulatory Cooperation and res_e/booksp_e/trade_in_services_and_development_
Standardization Policies (WP.6), Geneva: UNECE. Available [Link]
at [Link]
World Customs Organization (WCO) and World Trade
United Nations Educational, Scientific and Cultural Organization (WTO) (2022), "WCO/WTO Study Report
Organization (UNESCO) (2021), Recommendation on the on Disruptive Technologies", WCO-WTO Paper, Geneva:
Ethics of Artificial Intelligence, Paris: UNESCO. Available at WTO. Available at [Link]
[Link] sp_e/wco-wto_e.pdf

United Nations Educational, Scientific and Cultural World Economic Forum (WEF) (2018), "Harnessing Artificial
Organization (UNESCO) (2023a), Readiness Assessment intelligence for the Earth", World Economic Forum Paper,

112
 BIBLIOGRAPHY

Geneva: World Economic Forum (WEF). Available at World Trade Organization (WTO) (2017), 20 Years of the
[Link] Information Technology Agreement, Geneva: WTO. Avail-
Intelligence_for_the_Earth_report_2018.pdf able at [Link]
20years_2017_full_e.pdf.
World Economic Forum (WEF) (2020), "A Roadmap for
Cross-Border Data Flows: Future-Proofing Readiness and World Trade Organization (WTO) (2018), World Trade Re-
Cooperation in the New Data Economy", World Economic port 2018, Geneva: WTO. Available at [Link]
Forum Paper, Geneva: World Economic Forum (WEF). english/res_e/publications_e/world_trade_report18_e.pdf.
Available at [Link]
ap-for-crossborder-data-flows-future-proofing-readiness- World Trade Organization (WTO) (2020a), World Trade Report
and-cooperation-in-the-new-data-economy/ 2020, Geneva: WTO. Available at [Link]
lish/res_e/booksp_e/wtr20_e/wtr20_e.pdf
World Economic Forum (WEF) (2022), "Unlocking Value
from Artificial Intelligence in Manufacturing", World World Trade Organization (WTO) (2020b), WTO TBT
Economic Forum Paper, Geneva: World Economic Forum Committee and regulatory measures: Prevention, not
(WEF). Available at [Link] litigation, Geneva: WTO. Available at [Link]
AI_in_Manufacturing_2022.pdf english/tratop_e/tbt_e/tbt_t40_20920_e.htm

World Economic Forum (WEF) (2023), "Global Risks Report World Trade Organization (WTO) (2021), Technical barriers
2023", World Economic Forum Insight Paper, Geneva: World to trade (Third eds.), Geneva: WTO.
Economic Forum (WEF). Available at [Link]
org/publications/global-risks-report-2023/ World Trade Organization (WTO) (2022), Thematic Session
on Digital Solutions for Conformity Assessment Procedures,
World Economic Forum (WEF) (2024a), "ChatWTO: An Geneva: WTO. Available at [Link]
Analysis of Generative Artificial Intelligence and International tratop_e/tbt_e/tbtts_e/tbtts080322pm_e.htm
Trade", World Economic Forum White Paper, Geneva; World
Economic Forum (WEF). Available at WEF_An_Analysis_ World Trade Organization (WTO) (2023a), Thematic Session
of_Generative_Artificial_Intelligence_and_International_ on Conformity assessment and e-commerce, Geneva: WTO.
Trade_2024.pdf ([Link]) Available at [Link]
tbt_0711202315_e/tbt_0711202315_e.htm
World Economic Forum (WEF) (2024b), "Global Risks
Report 2024", World Economic Forum Insight Report, World Trade Organization (WTO) (2023b), Thematic Session
Geneva: World Economic Forum (WEF). Available at on Regulatory Cooperation between Members on
[Link] Cybersecurity, Geneva: WTO. Available at [Link]
port-2024/in-full/global-risks-2034-over-the-limit/ [Link]/english/tratop_e/tbt_e/tbt_2006202315_e/
tbt_2006202315_e.htm
World Health Organization (WHO) (2021), "Ethics and
governance of artificial intelligence for health: WHO World Trade Organization (WTO) (2023c), Thematic
guidance", World Health Organization Paper, Geneva: World Session on Regulatory Cooperation between Members on
Health Organization (WHO). Available at [Link] Intangible Digital Products, Geneva: WTO. Available at
bitstream/handle/10665/341996/9789240029200-eng. h t t p s : / / w w w. w t o . o r g / e n g l i s h / t r a t o p _ e / t b t _ e /
pdf?sequence=1 tbt_2006202310_e/tbt_2006202310_e.htm

World Intellectual Property Organization (WIPO) (2020), World Trade Organization (WTO) (2023d), Thematic session
"WIPO conversation on Intellectual Property (IP) and on the Use of Digital Technologies and Tools in Good
Artificial Intelligence (AI)", WIPO Paper No. WIPO/IP/ Regulatory Practices, Geneva: WTO. Available at https://
AI/2/GE/20/1 REV., Geneva: World Intellectual Property [Link]/english/tratop_e/tbt_e/tbt_0711202310_e/
Organization (WIPO). Available at [Link] tbt_0711202310_e.htm
edocs/mdocs/mdocs/en/wipo_ip_ai_2_ge_20/wipo_ip_
ai_2_ge_20_1_rev.pdf World Trade Organization (WTO) (2024a), Disputes by
agreement, Geneva: WTO. Available at [Link]
World Intellectual Property Organization (WIPO) (2024), org/english/tratop_e/dispu_e/dispu_agreements_index_
"Getting the Innovation Ecosystem Ready for AI", WIPO [Link]?id=A22#selected_agreement
Paper, Geneva: World Intellectual Property Organization
(WIPO). Available at [Link] World Trade Organization (WTO) (2024b), "Guidelines on
en/wipo-pub-2003-en-getting-the-innovation-ecosystem- conformity assessment procedures", WTO Committee on
[Link] Technical Barriers to Trade Decision No. G/TBT/54, Geneva:
WTO. Available at [Link]
World Trade Organization (WTO) (2004), A handbook on [Link]?filename=q:/G/TBT/[Link]&Open=True
the WTO dispute settlement system (eds.), Cambridge:
Cambridge University Press. World Trade Organization (WTO) (2024c), Services
Domestic Regulation, Geneva: WTO. Available at https://
113
 BIBLIOGRAPHY

[Link]/english/tratop_e/serv_e/sdr_factsheet_ Microsoft Source. Available at: [Link]

feb24_e.pdf source/asia/features/with-help-from-next-generation-ai-
indian-villagers-gain-easier-access-to-government-services/
World Trade Organization (WTO), Organisation for Economic
Co-operation and Development (OECD), International Zhou, B. (2019), Artificial Intelligence and Copyright
Monetary Fund (IMF) and United Nations (UN) (2023), Protection -Judicial Practice in Chinese Courts, WIPO.
Handbook on Measuring Digital Trade (Second eds.), Available at [Link]
Washington DC: International Monetary Fund. Available at ip/en/artificial_intelligence/conversation_ip_ai/pdf/ms_
[Link] china_1_en.pdf

Yee, C. (2023). “With help from next-generation AI, Indian

villagers gain easier access to government services”.

114
 World Trade Organization
154, rue de Lausanne
CH-1211 Geneva 2
Switzerland
Tel: +41 (0)22 739 51 11
[Link]

WTO Publications
Email: publications@[Link]
WTO Online Bookshop
[Link]

Report designed by Touchline

Printed by the World Trade Organization.

© World Trade Organization 2024.
Print ISBN 978-92-870-7661-8
Web ISBN 978-92-870-7660-1
Published by the World Trade Organization.
Artificial intelligence (AI) is transforming the way we live, work, produce and trade.
As it further develops, AI is expected to unlock unprecedented economic and societal
opportunities. However, it is also a source of significant risks and challenges.

This report examines the intersection of AI and international trade. It discusses how AI may
shape the future of international trade by reducing trade costs, improving productivity and
expanding economies' comparative advantages. The report reviews some key trade policy
considerations, in particular the urgent need to address the growing AI divide between
economies and between large and small firms, as well as data governance and
intellectual property issues. It examines how to guarantee the trustworthiness of AI
without hindering trade. The report also provides an overview of domestic, regional and
international government initiatives to promote and regulate AI, and highlights the resulting risk
of regulatory fragmentation.

Finally, the report discusses the critical role of the WTO in facilitating AI-related trade,
ensuring trustworthy AI and addressing emerging trade tensions, noting that the rapid
evolution of AI is prompting questions about the implications of AI for international
trade rules.