DEPARTMENT OF INFORMATION TECHNOLOGY

GE3791 – Human Values And Ethics

ASSIGNMENT II – REVIEW PAPER

Submitted by
Rohan Sharma P (812022205040)
Sabhareesan G (812022205041)
Sidharraj V (812022205047)

M.A.M. COLLEGE OF ENGINEERING AND TECHNOLOGY

TIRUCHIRAPPALLI – 621105

ANNA UNIVERSITY: CHENNAI 600025

SEPTEMBER 2025
 Marks Split up

Marks Allotted Rohan Sharma P Sabhareesan G Sidharraj V

Rubrics RBTL
(40)
Technical
K2 12
Comprehension of
Reference papers.
Critical Thinking K3 10
Sentence Structure, clarity of 10
writing, and Presentation K4
Reference papers citation K2 4
Timely Submission K3 4

Total Marks 40

[Link] COs RBTL Marks(40) Rohan Sharma P Sabhareesan G Sidharraj V

1. CO701.1 K2 13
2. CO701.2 K2 13
3. CO701.3 K3 04
4. CO701.4 K3 05
5. CO701.5 K3 05
Total Marks
40

Faculty in Charge
 2023 27th International Computer Science and Engineering Conference (ICSEC)

Ethics in AI Governance: Comparative

Analysis, Implication, and Policy
Recommendations for the Philippines
Angelis O. Arcilla Jr.1, Adrian Kim V. Espallardo1, Clio Andrei J. Gomez1, Edrian Miles P. Viado1, Vince Jeremy T.
Ladion1, Ruben Andrei T. Naanep 1, Aaron Raphael L. Pascual1, Edcel B. Artificio2, and Orland D. Tubola3
1
Department of Computer Engineering, Polytechnic University of the Philippines
2
Faculty of Computer Engineering Department, Polytechnic University of the Philippines
3
Research Institute for Strategic Foresight and Innovation, Polytechnic University of the Philippines
2023 27th International Computer Science and Engineering Conference (ICSEC) | 979-8-3503-4210-9/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICSEC59635.2023.10329756

Abstract— This study provides a comparative analysis of Furthermore, the absence of adequate frameworks has
Artificial Intelligence (AI) ethical guidelines by looking at six contributed to AI's violation of human rights. Frameworks
different sources from the USA, European Union, India, must be established to facilitate the retooling or upskilling of
Philippines, UNESCO, and CAIDP. The researchers extracted government agencies responsible for overseeing AI
key terms from these guidelines and categorized them based on governance and regulation [7]. Meanwhile, in sustainability
their frequency across the six sources to get a deeper management and decision-making, the utilization of AI has
understanding of their content and identify central themes and been limited. Previous sustainability management and
topics across each guideline. In doing so, the paper was able to decision-making studies have primarily relied on standard
identify ethical considerations and challenges of implementing
supervised learning methods, focusing on single dimensions
AI in the Philippine context, including data collection, use and
sharing, and AI development. The insights from the literature
such as land use or climate [8]. However, this approach needs
and document reviews and the analysis of the extracted key to capture the complexity of real-world scenarios.
terms were used to craft considerations in formulating AI policy To mitigate these risks, it is crucial to prioritize ethical
for the Philippines. These include (1) defining ethical principles considerations in developing and implementing AI systems,
and values in developing and implementing AI, (2) establishing implementing robust safeguards [9]. Responsible AI
mechanisms for ensuring fairness and preventing bias, (3) implementation requires ethical guidelines and data policies
implementing mechanisms to ensure security and mitigating
aligned with international standards on data protection and
risks in AI systems, and (4) assessing potential impacts on AI
technologies. Overall, the study uncovered that existing
responsible data practices [10]. Five key ethical principles—
guidelines and legal frameworks on AI ethics seem to have put justice, transparency, fairness, nonmaleficence, responsibility,
less emphasis on ethics and fairness. This underpins the need for and privacy—have been identified globally, but consensus on
the PH to ensure that its localized AI ethics policy is clear about interpretation and implementation remains challenging.
its definition of what is ethical and fair use of AI across sectors, Contextual factors like culture, society, and legal frameworks
industries, demographics, and contexts. influence the interpretation and implementation of ethical AI
principles across different countries, making flexibility and
Keywords—Artificial Intelligence (AI), AI Ethics, Data adaptability crucial in developing AI governance frameworks
Policy, AI Governance, Guidelines tailored to specific needs, such as those in the Philippines [11].
In line with these considerations, the research aims to
I. INTRODUCTION conduct a comparative analysis of AI ethics and governance
frameworks in various international organizations and
Artificial Intelligence (AI) has evolved significantly over
countries, including the United Nations Educational,
the past decade, transforming various industries and creating
Scientific and Cultural Organization (UNESCO), and Center
new possibilities for people worldwide [1]. While AI
for AI and Digital Policy (CAIDP), United States, the
technology has expanded opportunities and improved lives, it
European Union, and India. Its objective is to analyze these
has also raised ethical concerns that must be addressed. The
areas' existing guidelines and practices and assess their
deployment of AI systems presents significant risks,
relevance to the Philippine context. Valuable insights gained
infringing human rights, which can disproportionately affect
from comparing these frameworks will inform the
marginalized groups [2-3].
development of tailored AI governance frameworks that meet
AI has had detrimental effects on human rights through the specific needs and requirements of the Philippines.
various means. Firstly, it has led to the exploitation of data.
Furthermore, the study will investigate AI ethics in the
Secondly, it has facilitated the reidentification and
Philippines, summarizing current policies and guidelines. It
deanonymization of individuals [4]. Thirdly, concerns have
will also go beyond the existing frameworks by identifying
arisen regarding bias and discrimination [5]. Fourthly, there
ethical considerations and challenges within the Philippine
needs to be sufficient diligence in upholding ethical practices.
context for data collection, use, and sharing, as well as AI
Lastly, the transparency of AI systems has been compromised,
development and other related areas. Through a thorough
amplifying the risks posed to human rights [6].
analysis of these ethical considerations and their implications,
the study aims to offer policy recommendations that promote

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on December 12,2023 at 07:31:23 UTC from IEEE Xplore. Restrictions apply.
979-8-3503-4210-9/23/$31.00 ©2023 IEEE 319
 2023 27th International Computer Science and Engineering Conference (ICSEC)

the responsible and equitable deployment of AI in the i) Data Privacy: This term refers to an individual's
Philippines. By fostering trust and ensuring the well-being of ability to manage personal information, including
individuals and society, this research seeks to contribute to how it is collected, used, and shared. Data privacy
advancing AI technologies while upholding ethical standards policies aim to protect sensitive data and ensure
within the Philippine context. compliance with relevant regulations, such as the
GDPR in European Union.
II. LITERATURE REVIEW ii) Data Protection: It involves safeguarding data
This literature review provides an overview of AI ethics and against unauthorized access, loss, alteration, or
data policy, exploring their definitions, concepts, and destruction. Data protection policies include
importance. It discusses key aspects of AI ethics, such as encryption, access controls, and data backup to
privacy, fairness, transparency, accountability, security, and mitigate risks and ensure data integrity.
social impact, along with the role of data policy, focusing on iii) Data Sharing: Parties should explicitly state what
the General Data Protection Regulation (GDRP). The review
data can be shared. Any data-sharing policy should
also examines UNESCO's "Recommendation on the Ethics of
Artificial Intelligence" and organizations like CAIDP's efforts make the shared data anonymous to protect data
to address the connection between AI and human rights. privacy.
Additionally, it explores the progress made in implementing iv) Consent and Opt-Out: Parties should obtain consent
AI ethics in countries such as the United States, the European from individuals before collecting and processing
Union, and India while highlighting the current state of data their data. Also, they should be able to opt out of
policy and AI governance in the Philippines. certain data collection practices or withdraw consent
at any time.
A. Definition and Concepts of AI Ethics and Data Policy AI Ethics and Data Policy share common objectives
centered around protecting individuals and establishing
AI Ethics involves establishing values, guidelines, and standards in the digital era. Both disciplines prioritize
standards for creating, implementing, and using AI protecting people's rights, privacy, and well-being in the
technologies [12]. It includes considerations such as privacy, increasingly data-driven and AI-powered world. By aligning
fairness, transparency, accountability, security, and the AI ethics with robust data policies, society can ensure that AI
broader impact of AI on society [13-18]. technologies are developed and deployed responsibly and
i) Privacy: It involves respecting privacy rights and ethically, thus fostering trust, fairness, and accountability.
complying with data protection regulations, with
measures in place to safeguard personal data [13]. B. UNESCO’s Recommendations on the Ethics of AI
ii) Fairness: It involves developing and implementing Adapting principles and concepts for AI ethics should
AI systems non-discriminately, avoiding bias based have an internationally recognised standard. In November
on characteristics such as race, gender, age, or 2021, UNESCO adopted the "Recommendation on the Ethics
socioeconomic status [14]. of Artificial Intelligence," marking a significant milestone in
developing global standards for AI ethics. Supported by all
iii) Transparency: This means that AI algorithms and
193 Member States, this recommendation serves as a
systems should be explainable and understandable,
normative framework to address ethical concerns related to AI
with users and stakeholders having access to and foster trustworthiness throughout the AI system life cycle.
information about how AI decisions are made, and It places transparency, fairness, and the protection of human
the data used to train AI models [15] rights and dignity at its core [2].
iv) Accountability: It involves establishing clear
responsibility and mechanisms for addressing harms UNESCO's recommendation contains comprehensive
caused by AI [16]. policy action areas that empower policymakers to translate
v) Security: It involves having robust security measures fundamental values and principles into actionable measures.
to prevent unauthorized access or misuse that could These areas include data governance, environmental
preservation, gender equality, education and research, and
compromise the integrity of AI systems or cause
health and social welfare. It promotes respect for human rights
harm [17].
and fundamental freedoms, advocates for environmental
vi) Social Impact: It involves considering the societal flourishing, emphasizes diversity and inclusiveness and aims
implications of AI deployment, including potential to create peaceful, just, and interconnected societies.
effects on employment, social inequality, and human
rights, with efforts to mitigate negative impacts and Moreover, the Recommendation outlines multiple
maximize positive outcomes [18]. principles that guide ethical AI practices. These principles
Data policy plays a pivotal role in developing AI ethics include avoiding harm, fairness and equal treatment, safety
due to the extensive use of data in training AI systems, making and security, sustainability, privacy and data protection,
regulations necessary. The General Data Protection human oversight and decision-making, responsibility and
Regulation (GDPR), a data policy implemented by the EU, accountability, transparency and explainability, awareness
establishes rules and standards to promote responsible data and literacy, stakeholder involvement and adaptable
handling, ensure privacy, and uphold data security [19]. The governance.
following concepts about data policy have been derived from By adopting this global standard, UNESCO aims to guide
the GDPR: and support its Member States in implementing and reporting
on the ethical recommendations. It seeks to address the
challenges and risks associated with AI while maximizing its

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on December 12,2023 at 07:31:23 UTC from IEEE Xplore. Restrictions apply.
320
 2023 27th International Computer Science and Engineering Conference (ICSEC)

benefits for society. The Recommendation emphasizes the to the publication of several ethical guidelines for robots and
need for ethical considerations to be integrated into AI AI, the emergence of new ethical standards, and the creation
development and deployment, promoting a human-centered of national AI strategies and advisory bodies [21]. Several
and rights-based approach to ensure the responsible and countries and regions, including the United States, European
inclusive use of AI technologies. Union, and India, have significantly progressed in
implementing AI ethics.
C. CAIDP on the Ethics of AI and Human Rights This synthesis examines their progress and initiatives related
In parallel, the Center for AI and Digital Policy (CAIDP) to the implementation of AI ethics:
emphasized addressing the connection between AI and
i) United States: US released a set of AI principles in
human rights. CAIDP, a non-profit organization, is
2016 [22], and has been working on developing
committed to ensuring that advancements in AI contribute to
ethical guidelines for AI technology. The Department
a more equitable and fair society. It advocates for a world of Defense (DoD) has established AI ethical standards
where technological advancements are made harmoniously covering five key areas: Responsibility, Equity,
with respect for human rights, rule of law, and democratic Traceability, Reliability, and Governability [23 - 24].
institutions [20]. Among the recommendations by
organizations and experts are: ii) European Union: The European Union has released
its guidelines for creating ethical and trustworthy AI
i) Establish Clear Prohibitions on AI Technology: In systems [25]. These guidelines outline seven key
the General AI guidelines, there are explicit principles that should be followed: ensuring human
prohibitions on the use of AI for batch monitoring oversight and agency, maintaining technical safety
and unit ratings, also known as "social ratings". and robustness, protecting the privacy and properly
UNESCO's ethical recommendations for AI adopted managing data, being transparent in operations,
these two bans. The European Parliament has promoting diversity and preventing discrimination
proposed several AI bans in the EU AI law, including and unfairness, supporting environmental and
social ratings, predictive policies, biometric societal well-being, and being accountable for
classification, and emotion detection. actions taken. Furthermore, the EU has proposed the
ii) Access, Inclusion, and Equity: Emphasizing the need AI Act [26, 27], which regulates AI systems across
to ensure that AI technologies are accessible, various domains and establishes a harmonized
inclusive, and do not worsen existing inequalities. framework for AI ethics.
This involves addressing biases, promoting diversity iii) India: In 2021, the National Institution for
in AI development, and considering the needs of Transforming India (NITI) published a document
marginalized communities. establishing ethical principles for the design,
iii) Mandate Human Rights Impact Assessments for development, and deployment of AI. This acts as a
High-risk AI Systems: The statement of CAIDP for guide for the AI ecosystem, promoting the
the United States pointed out the need for mandated responsible implementation of AI [28].
human rights impact assessment before deploying iv) Other Countries: The Organization for Economic
any AI systems. The United States relies on Co-operation and Development (OECD) has
guidelines that were released decades ago. The developed a set of AI principles that provide a
CAIDP urges the US to publish new guidelines and framework for governments and businesses [29].
be more accountable for the risks and negative
impacts the AI systems pose. Providing impact E. Current State of Data Policy and AI Governance in the
assessments before deployment will also avoid Philippines
failures and malfunctions leading to certain Despite the progress made in AI ethics implementation by
complications. different organizations and countries, challenges persist in
iv) Make transparency Meaningful: In human rights, AI the Philippines. One significant challenge is the need for
transparency is essential. A decision and action made more consensus on the definition of ethical AI, as different
by an AI should always be able to be countered or perspectives and interpretations exist [30]. This lack of
stopped by individuals. Otherwise, the AI system
consensus hampers the development of clear and
should not be deployed and continued.
As the progress of AI technology unfolds, it presents both comprehensive ethical guidelines that can effectively guide
promising possibilities and significant challenges concerning AI development and deployment in the country. Additionally,
human rights. The Center for AI and Digital Policy (CAIDP), the rapid pace of technological development poses challenges
a dedicated non-profit organization, stands at the forefront of in keeping up with the ethical implications of AI
addressing the critical connection between AI and human advancements, further complicating the establishment of
rights. Their unwavering commitment ensures that AI robust ethical frameworks.
advancements contribute to a more equitable and fair society.
Despite this, in May 2021, the Department of Trade and
D. AI Ethics Implementation in Other Countries Industry (DTI) introduced the National Artificial Intelligence
Due to the active involvement of UNESCO and CAIDP, Strategy for the Philippines, positioning the Philippines as
their contribution has established comprehensive frameworks one of the initial 50 nations to adopt a comprehensive national
covering complex AI implications on human rights, the rule policy and guidelines concerning AI [31]. However, the
of law, and democratic principles. The collective effort has led country's AI implementation lags behind global benchmarks.

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on December 12,2023 at 07:31:23 UTC from IEEE Xplore. Restrictions apply.
321
 2023 27th International Computer Science and Engineering Conference (ICSEC)

According to a study on AI investment in Southeast Asia, the To gather the necessary data, the researchers relied on
Philippines is behind developed nations by 2 to 3 years per secondary sources. These included reviews of AI governance
capita AI investment, with a value of only 0.01 US dollars. policies and ethical guidelines from countries such as the
Nevertheless, integrating emerging AI technologies holds the United States, the European Union, India, and the Philippines.
This study also considered guidelines provided by
potential to contribute 92 billion dollars to the Philippine organizations like UNESCO and CAIDP. These sources were
economy by 2030 [32]. To ensure the continuous crucial for comprehensively understanding the diverse
advancement of AI in the country, the Philippines must perspectives and approaches to AI governance and ethics.
establish appropriate ethical guidelines for AI developments.
To collect the secondary data, the researchers employed
The AI Governance Framework report in the Philippines multiple methods. This involved accessing government
examines the fundamental principles that guide the design, websites and relevant agencies to acquire copies of their AI
creation, implementation, and utilization of AI within the governance policies and ethical guidelines. Additionally, the
study gathered secondary data from government reports,
country. It incorporates seven principles derived from the
academic articles, and news articles focusing on AI
Organization for Economic Co-operation and Development governance and ethics in the selected countries. By combining
(OECD) 's Recommendations in AI. These principles information from these sources, the researchers were able to
incorporate inclusive growth, sustainable development, and develop a comprehensive dataset for our comparative
the promotion of well-being; values and fairness that analysis.
prioritize humans; transparency and the ability to be By expanding the research beyond a single country and
explained; robustness, security, and safety; accountability; incorporating insights from different regions and
trustworthiness; and the protection of privacy [32]. organizations, our study aimed to provide a broader
perspective on AI governance and ethics. Through this
In May 2023, House Bill No. 7913, Artificial Intelligence examination, the researchers aimed to contribute to the
(AI) Regulation Act, was introduced in the House of responsible and ethical deployment of AI technologies in the
Representatives of the Philippines during the Nineteenth Philippines.
Congress, First Regular Session [33]. The bill's objective is
to establish a regulatory framework for developing, applying, B. Resource Materials
and using AI systems within the country. It provides six AI ethical principles from the USA, EU, India,
guiding principles: sustainable development, inclusive Philippines, UNESCO, and CAIDP were selected as targets
growth, and well-being; human-centered values; robust for content analysis. The following shows a brief overview
security and safety; accountability; transparency and these of AI ethical guidelines:
explainability; and trust. Also, it proposes the creation of the i) Artificial Intelligence and Human Rights [20]: AI
Philippine Council on AI, which will oversee the Guidelines from CAIDP (2023)
implementation of AI regulations and coordinate related
ii) Guidance for Regulation of Artificial Intelligence
efforts.
Applications [23]: AI Guidelines from USA (2020)
The bill outlines specific roles and responsibilities for iii) Responsible AI: Approach Document for India [28]:
various government agencies and defines prohibited acts AI Guidelines from India (2021)
associated with AI. The AI Regulation Act is designed to
promote and regulate the development and utilization of AI iv) UNESCO’s Recommendations on Ethics of AI [2]:
in the Philippines. The Act recognizes the significance of AI Guidelines from UNESCO (2019)
science and technology for national progress and seeks to v) Ethics Guidelines for Trustworthy AI [25]: AI
leverage AI's potential to enhance the lives of Filipinos, local Guidelines from European Union (2019)
industries, and the overall economy. vi) House Bill No. 7913 [33]: Artificial Intelligence
To propel the country's AI industry, it is imperative to Regulation Act (2023)
acknowledge the need for a comprehensive AI policy to
advance the industry further. However, the current state of AI C. Research Tools
governance in the Philippines is still in its early stages of To conduct the study, the proponents utilized Key Term
development and requires improvements to align with global Extraction [34], a multilingual system that identifies key terms
standards. from digital documents. This research tool is specifically
designed to automatically extract keywords, a crucial
technology used in advanced information retrieval systems.
III. RESEARCH METHODOLOGY By utilizing Key Term Extraction, enhances the analysis by
efficiently identifying significant keywords and visually
A. Research Design and Approach representing their prominence. This method can help
The research design will discuss and examine the policies contribute to a deeper understanding of the content and allow
of different countries and organizations. The objective was to for the rapid identification of central themes and topics within
identify the similarities and differences in how these policies the text data.
are formulated and the ethical considerations they address.

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on December 12,2023 at 07:31:23 UTC from IEEE Xplore. Restrictions apply.
322
 2023 27th International Computer Science and Engineering Conference (ICSEC)

V. IMPLICATION IN THE PHILIPPINES of ethical AI use. This could include initiatives in

schools, universities, and professional training
The section discusses the comparison between AI
programs to equip individuals with the necessary
guidelines in the Philippines and global standards. Also, the
knowledge and skills to navigate AI technologies
challenges and opportunities for AI ethics and data policy responsibly.
implementation in the Philippines are then examined,
focusing on education and awareness, data privacy rights, ii) Data Privacy Rights: A crucial requirement is data for
algorithm discrimination, and collaboration in AI AI systems to continue evolving and advancing.
governance. Lastly, recommendations for improving the AI Many companies collect user data in exchange for
guidelines in the Philippines include defining ethical their services. However, some companies have been
principles, ensuring fairness, enhancing security, and known to exploit this data to manipulate individuals
assessing societal and environmental impacts. or to use this data to train their algorithms. An
example of a prominent case demonstrating
mismanagement is the Facebook data breach, where
A. Philippines vs Global AI Guidelines
millions of users' personal information was breached
Although AI guidelines in the Philippines are still in their and disclosed to Cambridge Analytica without their
infancy, House Bill No. 7913 will pave the way for the future consent [36].
of the AI governance framework in the Philippines. However, To safeguard data privacy rights, a variety of
based on the results and discussion of this study, it is essential measures can be enacted. Companies must acquire
to note that the bill needs to include more emphasis on social explicit consent from users before gathering and
impact, security, fairness, and ethics. handling their personal data. This consent should be
well-informed and precise, explicitly stating the
UNESCO's Recommendations in the Ethics of AI pose
purpose of data collection and any involvement of
the most robust AI guidelines among all guidelines presented
third parties [15]. Furthermore, individuals should
in this study. Their recommendations have set the standard possess the right to access, rectify, and erase their
and served as a benchmark for developing other AI personal data [30].
guidelines. It recommends adapting principles for an ethical
framework that promotes the responsible development and Emphasizing the importance of safeguarding data
use of AI technologies. UNESCO's guidelines emphasize the privacy in the age of AI in the Philippines is essential.
importance of human rights, transparency, explainability, and Strict enforcement of the existing Data Privacy law
accountability in AI systems. [7] and careful consideration of the implementation of
the AI Council proposed by House Bill No. 7913,
In comparison, the Philippines' House Bill No. 7913 is a known as the "Artificial Intelligence (AI) Regulation
step towards AI governance, but it is still in its early stages. Act," are crucial steps to uphold and maintain order in
The bill primarily focuses on establishing the Philippine data privacy rights.
Artificial Intelligence Council and the AI Research and
iii) Algorithm Discrimination: Bias in AI algorithms is a
Development Program. While the bill acknowledges the significant concern that has garnered considerable
potential benefits of AI, it falls short in explicitly addressing attention in recent years. It refers to the unfair or
crucial aspects such as the ethical implications, fairness, discriminatory outcomes of using AI systems. These
potential biases, and the societal impact of AI. biases can stem from various factors, including the
data employed to train the algorithms, the design
B. Challenges and Opportunities for AI Ethics and Data decisions made during their development, or the
Policy Implementation in the Philippine Context inherent biases of the individuals involved [37].
Due to the dynamic and rapidly changing landscape of AI, The Philippines should emphasize the importance of
its ethical implementation has become a critical concern in addressing algorithmic discrimination by urging all
the Philippines. By this, there is a need to examine the stakeholders to ensure fairness in AI algorithms
progress and initiatives being undertaken to address AI ethics before deploying them. This objective can be
in the country. By doing so, the Philippines can identify areas accomplished through ethical considerations,
of improvement and implement measures to ensure technical measures, and regulatory frameworks.
responsible and ethical AI practices. Below addresses key
Developers and organizations should be encouraged
areas of concern: to evaluate and mitigate biases arising from the data
i) Education and Awareness: A significant obstacle lies used for training algorithms and the design choices
in need for more awareness and comprehension of AI made during their development. By actively
ethics and data policy among the public, promoting fairness and accountability, all parties
policymakers, and businesses. It is crucial to educate involved can collaborate towards eliminating
stakeholders about the potential advantages and risks discriminatory outcomes in AI systems.
associated with AI and emphasize the significance of
ethical considerations in its development and iv) Collaboration in AI Governance: Developing
utilization [35]. practical AI ethics and data policies requires
collaboration between government, industry, civil
The Philippines can invest in public awareness society organizations, and academia. Establishing
campaigns, educational programs, and training clear governance frameworks and multi-stakeholder
initiatives to promote AI literacy and foster a culture platforms can help facilitate dialogue, knowledge

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on December 12,2023 at 07:31:23 UTC from IEEE Xplore. Restrictions apply.
324
RIVIEW PAPERS
[1, 2, 3, 4, 5]
 ll
OPEN ACCESS

Article
Worldwide AI ethics: A review
of 200 guidelines and recommendations
for AI governance
Nicholas Kluge Correˆ a,1,2,7 Camila Galva˜ o,3 James William Santos,1,* Carolina Del Pino,4 Edson Pontes Pinto,3
Camila Barbosa,1 Diogo Massmann,1 Rodrigo Mambrini,3 Luiza Galva˜ o,5 Edmund Terem,6 and Nythamar de Oliveira1
1Graduate Program in Philosophy, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
2
Graduate Program in Philosophy, University of Bonn, Bonn, North Rhine-Westphalia, Germany
3Graduate Program in Law, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
4Psychology Undergraduate Program, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
5
Law Undergraduate Program, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
6Graduate Program in Philosophy, University of Johannesburg, Porto Alegre, Johannesburg, South Africa
7Lead contact

*Correspondence: [Link]@[Link]
[Link]

THE BIGGER PICTURE After the AI winter in the late 80s, AI research has experienced remarkable growth.
Currently, a lot of work is taking place to define the values and ideas that should guide AI advances. A key
challenge, however, lies in establishing a consensus on these values, given the diverse perspectives of
various stakeholders worldwide and the abstraction of normative discourse. Researchers and policy
makers need better tools to catalog and compare AI governance documents from around the world and
to identify points of divergence and commonality.

SUMMARY

The utilization of artificial intelligence (AI) applications has experienced tremendous growth in recent years,
bringing forth numerous benefits and conveniences. However, this expansion has also provoked ethical con-
cerns, such as privacy breaches, algorithmic discrimination, security and reliability issues, transparency, and
other unintended consequences. To determine whether a global consensus exists regarding the ethical prin-
ciples that should govern AI applications and to contribute to the formation of future regulations, this paper
conducts a meta-analysis of 200 governance policies and ethical guidelines for AI usage published by public
bodies, academic institutions, private companies, and civil society organizations worldwide. We identified at
least 17 resonating principles prevalent in the policies and guidelines of our dataset, released as an open
source database and tool. We present the limitations of performing a global-scale analysis study paired
with a critical analysis of our findings, presenting areas of consensus that should be incorporated into future
regulatory efforts.

INTRODUCTION puter science, the most frequently submitted sub-categories

for publications are "computer vision and pattern recognition,"
Since the last period of reduced interest in artificial intelligence "machine learning," and "computation and language," 1 i.e.,
(AI), the "AI winter" from 1987 to 1993, the field of AI research areas where machine learning (a sub-field of AI research) has es-
and industry has witnessed significant growth. This growth en- tablished itself as the reigning paradigm.
compasses various aspects, including the development of new Moreover, investment in AI-related companies and startups
technologies, increased investment, greater media attention, has reached unprecedented levels, with governments and ven-
and expanded capabilities of autonomous systems. A study ture capital firms investing over $90 billion (USD) in the United
analyzing the submission history on ArXiv from 2009 to 2021 re- States alone in 2021, accompanied by a surge in the registration
veals that computer science-related articles have become the of AI-related patents.2 While these money-field advancements
most prevalent type of material submitted, increasing 10-fold have brought numerous benefits, they also introduce risks and
starting in 2018. Furthermore, within the broad scope of com- side effects that have promoted several ethical concerns, like

Patterns 4, 100857, October 13, 2023 ª 2023 The Author(s). 1

This is an open access article under the CC BY license ([Link]
 ll
OPEN ACCESS Article
risks to user privacy, the potential for increased surveillance, the stitutions (e.g., AI Now Institute17), and professional associations
environmental cost of the industry, and the amplification of prej- (e.g., IEEE18), among other types of institutions, representing a
udices and discrimination on a large scale, which can dispropor- multi-stakeholder sample that for years was the most extensive
tionately harm vulnerable groups. Consequently, the expansion collection of AI guidelines analyzed systematically.
of the AI industry has given rise to what we refer to as the "AI One of the main findings in Jobin et al.’s3 work was the detec-
ethics boom," i.e., a period marked by an unprecedented de- tion of the most common ethical principles in the discourse of the
mand for regulation and normative guidance in this field. evaluated documents, those being transparency, justice/equity,
One of the central questions surrounding this boom is the non-maleficence, accountability, privacy, beneficence, freedom
determination of what ethical premises should guide the devel- and autonomy, trust, dignity, sustainability, and solidarity. Of
opment of AI technologies. And to answer this question, a these 11 ethical principles cited, five were the most recurrent:
plethora of principles and guidelines have been proposed by transparency (86%), justice (81%), non-maleficence (71%), re-
many stakeholders. However, the consensus and divergences sponsibility (71%), and privacy (56%). Furthermore, Jobin
in these varied discourses have yet to be extensively accessed. et al.’s work is careful not to impose any normative guidelines
For instance, do Silicon Valley-based companies follow the for the effectiveness of the mentioned principles. It attempts to
same precautions as major Chinese technology firms? Are these raise the issue and map the global picture in a pioneer work of
concerns relevant to end-users in countries with diverse cultural descriptive ethics. However, their limited sample, where, for
and social norms? Establishing a consensus to support the example, no Latin American countries are mentioned, makes
global regulations currently under discussion is of paramount the true global representativeness of the results questionable,
importance in both a practical and theoretical sense. a limitation recognized by Jobin et al. as one of their blind spots.
To address these questions, we draw inspiration from previ- Thilo Hagendorff19 conducted another study that presented a
ous works by meta-analysts and meticulously survey a wide similar type of analysis. In his research, Hagendorff focused on
array of available ethical guidelines related to AI development. a smaller sample of 21 documents. He excluded those older
These sources include governance policies of private com- than 5 years and that only addressed a national context unrelated
panies, academic institutions, and governmental and non- to AI, such as data science and robotics. Furthermore, Hagendorff
governmental organizations, as well as ethical guidelines for AI did not consider corporate policies but deliberately selected doc-
usage. By analyzing 200 documents in five different languages, uments deemed relevant in the international discourse (IEEE,
we gathered information on what ethical principles are the Google, Microsoft, and IBM) based on his evaluation criteria.
most popular, how they are described, where they come from, Even working with a smaller sample, Hagendorff’s findings
their intrinsic characteristics, and much else. Our primary goal corroborate with those of Jobin et al.,3 where the most mentioned
was to identify the most advocated ethical principles, to examine principles found were accountability (77%), privacy (77%), justice
their global distribution, and to assess if there is a consistent un- (77%), and transparency (68%). Hagendorff (like Jobin et al.) also
derstanding of these principles. Ultimately, this analysis aims to mentions the underrepresentation of institutions in South America,
determine whether a consensus exists regarding the normative Africa, and the Middle East as a clear bias of his sample.
discourse presented in ethical guidelines surrounding AI
Hagendorff is also more critical in his analysis, presenting
development.
concluding points such as the following:

State of the art d The lack of attention given to questions related to labor
One of the first studies to promote a meta-analysis of published rights, technological unemployment, the militarization of
AI ethical guidelines was that of Jobin et al.3 In this study, these AI, the spread of disinformation, and the misuse/dual-use
authors sought to investigate whether a global agreement on of AI technology
emerging questions related to AI ethics and governance would d The lack of gender diversity in the tech industry and AI
arise. The research identified 84 documents containing ethical ethics
guidelines for intelligent autonomous systems using the d The short, brief, and minimalist views some documents
Preferred Reporting Items for Systematic Reviews and Meta- give to normative principles
Analyses framework (originally developed for analysis in the d The lack of technical implementations for how to imple-
healthcare sector).4 At the time, some of them were one of the ment the defended principles in AI development
most cited guidelines in the literature, like the Organization for d The lack of discussion on long-term risks (e.g., artificial
Economic Co-operation and Development (OECD) Recommen- general intelligence safety, existential risks)
dation of the Council on Artificial Intelligence, 5 the High-Level
Expert Group on AI Ethics Guidelines for Trustworthy AI,6 the While the works of Hagendorf and Jobin et al. are valuable
University of Montre´ al Declaration for responsible development contributions to AI ethics, we question the criteria for filtering
of artificial intelligence,7 the Villani Mission’s French National documents used by these works. We argue that if we want to
Strategy for AI,8 among many others. investigate the consensus regarding the normative dispositions
Jobin et al.’s3 sample also contained documents from govern- of different countries and organizations regarding AI, we should
mental organizations (e.g., Australian Government Department not use popularity-based filtering. In other words, a descriptive
of Industry Innovation and Science 9), private companies (e.g., ethics evaluation should take as many viewpoints as possible if
SAP,10 Telefonica,11 IBM12), non-governmental organizations it aims to make a solid description of the "global landscape."
(e.g., Future Advocacy,13 AI4People14), non-profit organizations As a last mention, we would like to cite the work done by Fjeld
(e.g., Internet Society,15 Future of Life Institute16), academic in- et al.,20 one of the first to present documents from Latin America.

2 Patterns 4, 100857, October 13, 2023

 ll
Article OPEN ACCESS

In their study, Fjeld et al. worked with 36 samples produced by a gendorff and Fjeld et al.20 was not "user-friendly" or clear,
variety of institution types, but as with Hagendorff, 19 Fjeld et al. something that we tried to overcome in our work.
also excluded data science, robotics, and other AI-related (4) Released with an open source dataset, making our work
fields/applications. According to these authors, eight principles reproducible and extendable.
were the most commonly cited in their sample: fairness/non-
discrimination (present in 100% of the analyzed documents), pri- The focus of this study is guidelines related to the ethical use of
vacy (97%), accountability (97%), transparency/explainability AI technologies. We refer to "guidelines" as documents concep-
(94%), safety/security (81%), professional responsibility (78%), tualized as recommendations, policy frameworks, legal land-
human control of technology (69%), and promotion of human marks, codes of conduct, practical guides, tools, or AI principles
values (69%). for the use and development of this type of technology. The reso-
Like in the study of Jobin et al.,3 Fjeld et al.20 cite the variability in nating foundation for most of these documents is the presence of
how such principles are defined as an important element to be ad- a form of ‘‘principlism,’’24 i.e., the use of ethical principles to sup-
dressed. For example, in the 2018 version of the Chinese Artificial port normative claims.
Intelligence Standardization White Paper,21 the authors mention Now, deconstructing the expression "AI technologies," with
that AI can serve to obtain more information from the population, "AI," our scope of interest encompasses areas that inhabit the
even beyond the data that has been consented to (i.e., violation of multidisciplinary umbrella that is artificial intelligence research,
informed consent would not undermine the principle of privacy), such as statistical learning, data science, machine learning (ML),
while the Indian National Strategy for Artificial Intelligence Discus- logic programming/symbolic AI, optimization theory, robotics,
sion Paper (National Institution for Transforming India)22 argues software development/engineering, etc. And with the term "tech-
that its population must become massively aware so that they nologies," we refer to specific tools/techniques, applications, and
can effectively consent to the collection of personal information. services. Thus, the term refers to technologies for automating
We consider the delineation of diverging principles across docu- decision processes and mimicking intelligent/expert behavior.
ments to be one of the standout strengths of Fjeld et al.’s work, Unlike previous works,19,20 we included areas disregarded previ-
which warrants replication on a broader scale. ously (software development, data science, robotics).
There is more meta-analytical work that has been done in AI
ethics that we will not cite in depth. For a complete review of Sources
meta-analytical research on normative AI documents, we We used as sources for our sample two public repositories, the
recommend the work done by Schiff et al., 23 which cites many ‘‘AI Ethics Guidelines Global Inventory,’’ from AlgorithmWatch,
other important works. and the ‘‘Linking Artificial Intelligence Principles’’ (LAIP) guide-
Now, shifting the perspective from past works to our own, we lines. The AlgorithmWatch repository contained 167 documents,
argue that many of the mentioned analyses, following Jobin while the LAIP repository contained 90.
et al.’s3 work, suffer from a small sample size. While works like Initially, we checked for duplicate samples between both re-
the ones done by Hagendorff19 and Fjeld et al.20 show more positories. After disregarding them, we also scavenged for
diversified categories of typologies or sample sizes, their limited more documents through web search engines and web
sample may hinder the generalization of their results. Meanwhile, scraping, utilizing search strings such as "Artificial Intelligence
Jobin et al. do not present some features (e.g., an extensive Principles," "Artificial Intelligence Guidelines," "Artificial Intelli-
exemplification of how principles diverge) presented in other gence Framework," "Artificial Intelligence Ethics," "Robotics
works that used a smaller sample. Also, we would like to point Ethics," "Data Ethics," "Software Ethics," and "Artificial Intelli-
out that none of these studies released their dataset in a form gence Code of Conduct," among other related search strings.
that would allow the replication of their findings, which makes We limited our search to samples written/translated in one of
all of these studies (if one would not be willing to redo all the the five languages our team could cope with: English, Portu-
work) irreproducible. guese, French, German, and Spanish.
To cover these shortcomings, we propose an updated review We were able to collect in this manner 200 documents. Thus,
of AI guidelines and related literature: Worldwide AI Ethics. after defining our pool of samples, we initiated the data collection
stage. We divided this stage into two phases.
Methodology
From the gaps pointed out in the previous meta-analyses, in this First phase
study, we sought to present the following to the AI community. In phase one, members of our team received different quotas of
documents. Each team member was responsible for reading,
(1) A quantitatively larger and more diverse sample size, as translating when needed, and hand-coding pre-established fea-
done by Jobin et al.3 Our sample possesses 200 docu- tures. The first features looked for were the following:
ments originating from 37 countries, spread over six con-
tinents, in five different languages. d Institution responsible for producing the document
(2) Combined with a more granular typology of document d Country/world region of the institution
types, as done by Hagendorff.19 This typology allowed d Type of institution (e.g., academic, non-profit, govern-
an analysis beyond the mere quantitative regarding the mental, etc.)
content of these documents. d Year of publication
(3) Presented in an insightful data visualization framework. d Principles (as done by Fjeld et al.,20 we broke principles
We believe the data presentation done by authors like Ha- into themes of resonating discourse)

Patterns 4, 100857, October 13, 2023 3

 ll
OPEN ACCESS Article
d Principles description (i.e., the words used in a document d Self-regulation/voluntary self-commitment: this category
to define/support a given principle) is designed to encompass documents made by private
d Gender distribution among authors (inferred through a first organizations and other bodies that defend a form of
name automated analysis) self-regulation governed by the AI industry itself. It also en-
d Size of the document (i.e., word count) compasses voluntary self-commitment made by indepen-
dent organizations (NGOs, professional associations, etc.).
For the gender inference part, after removing documents with d Recommendation: this category is designed to encom-
unspecified authors, we performed a name-based gender anal-
pass documents that only suggest possible forms of
ysis. Given the variety/diversity that names can possess, it was governance and ethical principles that should guide orga-
necessary to use automation to infer gender encodes (male/fe- nizations seeking to use, develop, or regulate AI tech-
male). To make an accurate inference, we also extracted (in addi- nologies.
tion to each author’s name) the most likely nationality associated
with each name. For this, we used (in addition to the country/origin We defined these categories as mutually exclusive (the pres-
of each document) an API service that predicts the most likely na- ence of a feature excludes the other). The third type of document
tionality associated with a given name. Finally, we used another classification pertains to the normative strength of the proposed
API service to infer gender based on first name plus nationality. regulation mechanism. In this regard, we established two distinct
You can find the code for our implementation in this repository: categories, drawing upon the definitions provided by the ‘‘Inno-
[Link] vative and Trustworthy AI.’’25
Regarding how we defined principles, in the first phase, we es-
tablished a list of principles so our team could focus their search. d Legally non-binding guidelines: these documents propose
We based these on past works mentioned in the "related works" an approach that intertwines AI principles with recommen-
section: accessibility, accountability, auditability, beneficence/ ded practices for companies and other entities (i.e., soft
non-maleficence, dignity, diversity, freedom/autonomy, hu- law solutions).
man-centeredness, inclusion, intellectual property, justice/eq- d Legally binding horizontal regulations: these documents
uity, open source/fair competition, privacy, reliability, solidarity, propose an approach that focuses on regulating specific
sustainability, and transparency/explainability. uses of AI through legally binding horizontal regulations,
We also used this phase to determine categories/types as- such as mandatory requirements and prohibitions.
signed to each document in the second phase. These types
were determined by the following: We defined them as mutually inclusive. The final type relates to
the perceived impact scope that motivates the evaluated docu-
(1) The nature/content of the document ment. With impact scope, we mean the dangers and negative
(2) The type of regulation that the document proposes prospects regarding the use of AI that inspired the type of
(3) The normative strength of this regulation normative propositions used. For this, three final categories
(4) The impact scope that motivates the document were defined and also posed as mutually exclusive.

The first type relates to the nature/content of the document. d Short-termism: we designed this category to encompass
documents in which the scope of impact and preoccupa-
d Descriptive: descriptive documents take the effort of pre-
tion focus mainly on short-term problems, i.e., problems
senting factual definitions related to AI technologies. These we are facing with current AI technologies (e.g., algorithmic
definitions serve to contextualize "what we mean" when
discrimination, algorithmic opacity, privacy, legal account-
we talk about AI.
ability).
d Normative: normative documents present norms, ethical
d Long-termism: we designed this category to encompass
principles, recommendations, and imperative affirmations
documents in which the scope of impact and preoccupa-
about what such technologies should be used/devel-
tion focus mainly on long-term problems, i.e., problems
oped for.
we may come to face with future AI systems. Since such
d Practical: practical documents present development tools
technologies are not yet a reality, we can classify these
to implement ethical principles and norms, be they qualita- risks as hypothetical or, at best, uncertain (e.g., sentient
tive (e.g., self-assessment surveys) or quantitative (e.g.,
AI, misaligned AGI, super-intelligent AI, AI-related existen-
debiasing algorithms for ML models).
tial risks).
We defined these first three categories as mutually inclusive d Short-termism and long-termism: we designed this cate-
(documents may have all these features combined). The second gory to encompass documents in which the scope of
type relates to the form of regulation that the document impact is both short and long-term, i.e., they present a
proposes. "mid-term" scope of preoccupation. These documents
address issues related to the short-termism category while
d Government regulation: this category is designed to also pointing out the mid/long-term impacts of our current
encompass documents made by governmental institutions AI adoption (e.g., AI interfering in democratic processes,
to regulate the use and development of AI, strictly (legally autonomous weapons, existential risks, environmental
binding horizontal regulations) or softly (legally non-binding sustainability, labor displacement, and the need for updat-
guidelines). ing our educational systems).

4 Patterns 4, 100857, October 13, 2023

 ll
OPEN ACCESS Article

autonomy of human decision-making must be preserved parency of an organization" or "the transparency of an al-
during human-AI interactions, whether that choice is indi- gorithm." This set of principles is also related to the idea
vidual or the freedom to choose together, such as the invi- that such information should be understandable to nonex-
olability of democratic rights and values, also being linked perts and, when necessary, subject to be audited.
to technological self-sufficiency of nations/states. d Truthfulness: this principle upholds the idea that AI tech-
d Human formation/education: such principles defend that nologies must provide truthful information. It is also related
human formation and education must be prioritized in our to the idea that people should not be deceived when inter-
technological advances. AI technologies require consider- acting with AI systems.
able expertise to be produced and operated, and such
knowledge should be accessible to all. These 17 principles contemplate all of the normative discourse
d Human-centeredness/alignment: such principles advo- we could interpret. At the end of this second phase, all docu-
ments received 13 features: origin country, world region, institu-
cate that AI systems should be centered on and aligned
tion, institution type, year of publication, principles, principles
with human values. This principle is also used as a
"catch-all" category, many times being defined as a collec- definition, gender distribution, size, type I (nature/content),
tion of "principles that are valued by humans" (e.g., type II (form of regulation), type III (normative strength), type IV
freedom, privacy, non-discrimination, etc.). (impact scope), plus identifiers and attachments like document
d Intellectual property: this principle seeks to ground the title, abstract, document URL, and related documents.
property rights over AI products and their generated
outputs. Our tool
d Justice/equity/fairness/non-discrimination: this set of prin- We used all information obtained during the second phase
ciples upholds the idea of non-discrimination and bias miti- to create the dataset that feeds our visualization tool: an interac-
gation (discriminatory algorithmic biases AI systems can tive dashboard. We created this dashboard using the Power
be subject to). It defends that, regardless of the different BItool ([Link]
sensitive attributes that may characterize an individual, [Link]). We also developed a secondary dashboard
algorithmic treatment should happen "fairly." (open source) using the Dash library ([Link]
d Labor rights: labor rights are legal and human rights related [Link]/worldwide_AI-ethics/), an open source framework
to the labor relations between workers and employers. In for building data visualization interfaces.
AI ethics, this principle emphasizes that workers’ rights The main distinction between our tool and Hagendorff’s
should be preserved regardless of whether labor relations table19 and Fjeld et al.’s graphs20 lies in its interactivity and the
are being mediated/augmented by AI technologies. flexibility to combine various filters without being confined to
d Cooperation/fair competition/open source: this set of prin- preconfigured orderings. This feature allows researchers to uti-
ciples advocates different means by which joint actions lize the tool to examine and question specific characteristics pre-
can be established and cultivated between AI stakeholders sent in their regions, to identify trends and behaviors, or to
to achieve common goals. It also relates to the free and explore categories relevant to their research focus. It is worth
open exchange of valuable AI assets (e.g., data, knowl- noting that we were among the first to openly release our data-
edge, patent rights, human resources). set, making our work accessible and reproducible.
d Privacy: the idea of privacy can be defined as the individ- Another distinguishing feature of our tool is its ability to
ual’s right to "expose oneself voluntarily, and to the extent condense large amounts of information into a single visualization
desired, to the world." This principle is also related to data panel. Our choice for such a way of presenting our data was to
protection related-concepts such as data minimization, make it easier to interpret how certain features interact with
anonymity, informed consent, and others. others. While previous works demonstrate the statistical distri-
d Reliability/safety/security/trustworthiness: this set of prin- bution of certain features, ours allows the user to see how these
ciples upholds the idea that AI technologies should be reli- features vary and are interconnected.
able, in the sense that their use can be truly attested as safe
and robust, promoting user trust and better acceptance of Limitations
AI technologies. As in past works, this analysis also suffers from a small sample.
d Sustainability: this principle can be interpreted as a mani- Our work represents a mere fraction of what our true global
festation of "intergenerational justice," wherein the welfare landscape on this matter is. Some of the main limitations we
of future generations must be considered in AI develop- encountered during our work are the following.
ment. In AI ethics, sustainability pertains to the notion
that the advancement of AI technologies should be ap- (1) The limited scope of languages we were able to interpret
proached with an understanding of their enduring conse- represents a language bias, potentially excluding relevant
quences, encompassing factors such as environmental perspectives.
impact and the preservation and well-being of non-human (2) Publication bias is also a concern, as the focus on pub-
life. lished guidelines may overlook valuable insights from
d Transparency/explainability/auditability: this set of princi- ongoing discussions in other forms of media.
ples supports the idea that the use and development of (3) The "guideline" scope excludes the academic work being
AI technologies should be transparent for all interested done worldwide (i.e., we did not consider academic pa-
stakeholders. Transparency can be related to "the trans- pers on AI ethics).

6 Patterns 4, 100857, October 13, 2023

 International Journal of Computers
Maikel Leon [Link]

Artificial Intelligence: Evolution, Challenges, Future and Governance

MAIKEL LEON
Department of Business Technology
University of Miami
Miami, Florida, USA

Abstract: - Artificial Intelligence (AI) has advanced far beyond its early days of symbolic reasoning into
an era driven by deep neural networks and generative models. These techniques now power medical
diagnostics, financial risk assessment, autonomous vehicles, and mass-scale content generation. Alongside
these breakthroughs, concerns regarding data privacy, algorithmic bias, misinformation, and environmental
sustainability have grown more urgent. This paper traces the evolution of AI from hand-crafted rule systems
to large language models and generative architectures, examining ethical and societal implications, including
biases in training data and deepfake disinformation. We explain how the Massive Multitask Language
Understanding benchmark highlights the increasing depth of AI language capabilities. The discussion then pivots
to governance frameworks, focusing on audit mechanisms, embedded ethical considerations, and international
policy efforts to ensure fairness, transparency, and equitable access. We also explore ecological solutions, such
as energy-efficient hardware and carbon-neutral data centers. Future trends like neuromorphic computing, hybrid
AI, and quantum-based approaches are opportunities and challenges for responsible AI development. This paper
underscores the critical need for proactive, inclusive governance to align AI progress with societal well-being and
global sustainability.
Key-Words: - AI bias and ethics, ML transparency, Massive Multitask Language Understanding benchmark.
Received: April 19, 2024. Revised: January 7, 2025. Accepted: March 11, 2025. Published: May 8, 2025.

1 Introduction However, these tools require robust privacy

Artificial Intelligence (AI) has rapidly evolved from safeguards and oversight to ensure equitable
a niche research domain into a pervasive force that access.
touches virtually every sector, including healthcare, • Generative AI (GenAI) platforms can assist in
transportation, finance, education, and beyond. Over creating personalized tutoring content, although
the last decade, AI has demonstrated superhuman some educators worry about diminishing critical
performance in areas as diverse as complex strategy thinking skills if students over-rely on automated
board games and biomedical image analysis [1]. systems [5, 6].
Concurrently, the growth of generative modeling
techniques has enabled AI systems to produce Nevertheless, the dual-use nature of AI
realistic text, images, and multimedia content. complicates governance efforts. Large language
These developments hold immense promise, ranging models (LLMs) like GPT-4 democratize information
from the automation of routine tasks to advanced access while enabling malicious actors to produce
medical diagnoses and personalized educational spam or deepfake text at scale. Computer vision
platforms. Yet, they also provoke concerns regarding techniques enhance medical imaging but risk privacy
economic disruptions, data privacy, and large-scale violations if patient information is not adequately
misinformation campaigns. protected. A 2023 study found that 68 percent of AI
Several powerful AI tools have already found systems used in healthcare do not include sufficient
mainstream adoption: documentation for bias auditing, indicating serious
oversight gaps [7].
• Recommendation engines tailor content or Governments, corporations, and civil society
products to individual preferences, potentially organizations have advocated for robust frameworks
improving user satisfaction and raising questions integrating technical and ethical considerations. Yet,
about echo chambers and filter bubbles [2–4]. many high-profile statements on ethical AI remain
superficial, lacking rigorous guidelines for auditing,
• AI-driven solutions for analyzing large medical accountability, or global coordination. As AI
datasets offer faster, more accurate diagnoses. is integrated more extensively into core societal

ISSN: 2367-8895 81 Volume 10, 2025

 International Journal of Computers
Maikel Leon [Link]

infrastructure, this discrepancy between rapidly modern AI language capabilities. Section IV covers
advancing AI capabilities and limited oversight pressing ethical and societal implications. Section V
becomes alarming. Recognizing that responsible AI explores governance frameworks designed to anchor
entails more than just technical gains, stakeholders AI in moral principles. Section VI provides an
must address issues like algorithmic fairness, human overview of future AI trends, and Section VII outlines
oversight, transparency, and sustainable resource eco-solutions to address AI’s environmental impact.
usage from the outset. Section VIII focuses on global governance and equity
Moreover, global AI adoption is uneven. concerns, while Sections IX and X present expanded
Regions with strong digital infrastructure and ample conclusions and potential avenues for future research,
investments in R&D often drive AI breakthroughs, respectively.
while other parts of the world lag. This disparity
fosters concerns about an evolving AI divide, where 2 From Hand-Crafted Rules to
only certain nations and communities reap the
technology’s benefits. Addressing such inequities Generative Models
necessitates: In this section, we examine the historical development
of AI methods, moving from symbolic logic-based
• Investment in AI capacity building and training, expert systems to today’s dynamic, data-driven
particularly in underserved regions [8, 9]. architectures based on Machine Learning (ML).
We highlight how these shifts have impacted
• Policies that promote the inclusive distribution of
AI capabilities and the challenges related to
data and computational resources.
transparency, performance, and resource needs.
• International partnerships that facilitate
knowledge sharing and encourage ethical 2.1 Symbolic Reasoning: Expert Systems
AI adoption [10, 11]. In the mid-to-late 20th century, AI focused heavily
on symbolic reasoning. Researchers aimed to encode
Table 1 summarizes select AI milestones, domain knowledge as logical rules, theorizing
illustrating how the field evolved from symbolic that intelligent behavior would emerge from
logic systems to the deep learning (DL) and GenAI explicit rule sets. MYCIN (1976), an early
approaches that now dominate. expert system, diagnosed bacterial infections
using if-then statements. This rule-based approach
Table 1: Select AI Milestones Over Time provided clear decision paths, as each outcome
Year Milestone followed a well-defined logical statement. However,
symbolic AI proved fragile and slow to adapt. The
1956 Dartmouth Workshop knowledge engineering process required extensive
1976 MYCIN Expert System human curation, and systems often struggled with
1997 Deep Blue defeats Kasparov ambiguous or incomplete information. Projects like
2012 AlexNet advances DL CYC (1984) showcased how maintaining millions of
2020 GPT-3 few-shot language capabilities rules remained unwieldy by 2020 [12].
2023 GPT-4 excels on a variety of benchmarks Limitations of symbolic AI include:
• Poor handling of uncertainty or incomplete data,
This paper thoroughly assesses AI’s evolution
leading to brittle performance.
from symbolic reasoning to the versatile generative
architectures now prevalent. We then explore critical • Reliance on costly domain experts to craft and
challenges in AI governance—ethical, societal, and update rule bases.
environmental—and identify practical measures
such as algorithmic auditing, certification, human • Difficulty transferring knowledge across tasks
oversight, and inclusive governance models. By due to domain-specific logic.
analyzing near-term risks and highlighting emerging
areas, such as neuromorphic hardware and quantum 2.2 ML and Data-Driven Approaches
AI, we underscore the necessity of adopting ethical From the early 1990s onward, the field shifted toward
guardrails to secure widespread benefits. ML, where algorithms learn patterns directly from
Section II examines how AI methods progressed data rather than relying on meticulously crafted rules.
from symbolic systems to deep and generative models Decision trees, support vector machines (SVMs), and
to guide readers through the subsequent content. Bayesian models became popular for handwriting
Section III discusses the Massive Multitask Language recognition, spam detection, and speech-processing
Understanding benchmark, providing insight into tasks. Rapidly growing datasets and the rise

ISSN: 2367-8895 82 Volume 10, 2025

 International Journal of Computers
Maikel Leon [Link]

of more affordable computing power accelerated 2.5 Hardware Innovations

progress. MNIST (1998), with its 70,000 digit Hardware and cloud infrastructures have catalyzed
images, became a standard benchmark; advanced AI’s expansion. High-performance chips such as
SVM variants achieved up to 99.3 percent accuracy Google’s TPU v4 (2021) deliver 275 teraflops,
on this dataset. While these models often performed reducing training times for common models to mere
well, interpretability was not a priority. A 2008 minutes. Neuromorphic chips, like Intel’s Loihi
study reported that SVM-based credit scoring systems 2, simulate spiking neurons for energy-efficient
provided no clear explanations for why specific loan computation [20]. These progressions enable AI
applications were rejected, placing applicants at a to tackle tasks once deemed intractable, though
disadvantage [7]. This opacity would later prompt the benefits often cluster in regions with ample
calls for explainable and transparent approaches [13, computational resources, reinforcing concerns about
14]. global inequalities. Most AI patents, approximately
78%, originate in the United States and China,
suggesting limited participation from lower-income
2.3 DL Renaissance countries [20]. As the gap in AI capabilities widens,
The 2010s witnessed a DL renaissance, fueled by bridging access and expertise becomes more critical.
large neural networks that automatically extracted Table 2 summarizes the core methods and relative
features from raw inputs. Convolutional neural advantages of different AI paradigms, providing
networks (CNNs) demonstrated remarkable accuracy context for these trends.
in image recognition tasks, notably in the 2012
ImageNet competition with AlexNet. Recurrent
Table 2: Key AI Paradigms and Their Characteristics
architectures like LSTMs enabled breakthroughs
in language translation and time-series analysis. Parad. Methods Benefits/Drawbacks
Through these approaches, functions that once Symb. Rule-based Transparent/brittle
required highly specialized feature engineering ML SVMs, DTs Adaptive/often opaque
became tractable with the right network designs DL CNNs, RNNs High accuracy/complex
and sufficient data. Nevertheless, the computational GenAI GANs, LLMs New content/hallucinate
demands of deep models soared. GPT-3 (2020)
required an estimated 3.14e23 floating-point
operations, equivalent to running a GPU for 355
years [15]. This colossal resource usage raised 3 MMLU
questions about environmental sustainability and Massive Multitask Language Understanding
access, especially for researchers in less affluent (MMLU) is a comprehensive benchmark to evaluate
institutions or regions [16]. Specialized hardware, language models’ broad knowledge and reasoning
including GPUs, TPUs, and ASICs, emerged to capabilities. This mini paper provides a historical
accelerate training and mitigate these issues. background of MMLU’s development, the theoretical
foundations and design principles underlying its
construction, and a practical explanation of how the
2.4 Generative Models benchmark works, including dataset composition,
Since the late 2010s, two developments have defined evaluation methodology, and applications. We trace
AI research. First, transfer learning allows large the motivation for MMLU’s creation in response
pretrained models to adapt rapidly to new domains to rapid progress on earlier benchmarks, outline its
with minimal retraining data. Models like BERT structure spanning dozens of tasks across diverse
(2018) revolutionized natural language processing domains, and discuss its significance as a measure
by leveraging unsupervised pretraining on massive of general language understanding in modern AI
text corpora and then fine-tuning for specialized systems.
tasks. Second, GenAI has risen to prominence, Over the past few years, natural language
capable of producing realistic images, text, and processing (NLP) benchmarks have driven rapid
even complex simulations. Generative adversarial advances in language model performance. Early
networks (GANs) like StyleGAN2 synthesize highly evaluation suites such as the General Language
plausible human faces, while LLMs like GPT-4 Understanding Evaluation (GLUE) benchmark [21]
display sophisticated language skills. Yet, these and its more challenging successor SuperGLUE
generative models often lack a built-in mechanism for [22] played a crucial role in measuring progress.
verifying factual correctness, leading to phenomena However, by 2019–2020, leading models had already
such as hallucination, where plausible but false achieved near or above human-level performance
information is presented [13, 17–19]. on these benchmarks, indicating that they no

ISSN: 2367-8895 83 Volume 10, 2025

 International Journal of Computers
Maikel Leon [Link]

longer sufficiently discriminated the capabilities of task-specific fine-tuning, thereby assessing the
state-of-the-art models [22, 23]. This prompted the generalization power gained from pretraining.
search for new, more comprehensive tests of language
understanding. MMLU was thus motivated by the desire to
One response was the introduction of the Massive measure comprehensive language understanding,
Multitask Language Understanding (MMLU) testing not just linguistic prowess or shallow pattern
benchmark [24] proposed as a far-reaching challenge matching, but the extent to which models have learned
encompassing a wide range of subjects and difficulty factual and procedural knowledge across domains.
levels, intended to evaluate a model’s general Upon its release, MMLU was markedly more difficult
knowledge and reasoning abilities beyond the for models than previous benchmarks: Early tests
narrow scope of prior benchmarks. Unlike GLUE showed that many contemporary models performed
and SuperGLUE, which focus on linguistic and only at random-guess levels (around 25% accuracy)
commonsense reasoning tasks, MMLU covers on this benchmark, underlining the challenge it posed
a broad spectrum of academic and professional [24]. Even the largest GPT-3 model of that time
domains. The motivation behind MMLU was to achieved only about 43.9% accuracy on MMLU, well
create an enduring benchmark that would remain below an estimated expert human accuracy of roughly
challenging even as models continued to improve, 90% [24, 25]. This gap highlighted the headroom
thereby providing a more realistic assessment of a MMLU provided for future improvement.
model’s understanding of open-world knowledge.
3.2 Design Principles and Foundation
3.1 Historical Background and Motivation The design of MMLU rests on several key principles
By late 2019, NLP researchers observed a intended to align the benchmark with a theoretical
disconnect between benchmark performance and ideal of broad, multitask language understanding.
true language understanding. Models like BERT and At its core, MMLU is grounded in the concept
its variants quickly saturated GLUE, and even the of evaluating knowledge transfer and recall from
tougher SuperGLUE was nearly solved in a short pretraining: models are tested in a zero-shot or
time [22, 23]. These benchmarks, while useful, few-shot setting on tasks they have never been
covered a limited range of tasks (mostly short text explicitly trained on, simulating how a human
understanding and commonsense reasoning) and leverages general education to answer novel
thus did not capture many aspects of language questions [24]. This approach focuses on emergent
competence. For example, they did not extensively knowledge in language models, i.e., what the model
test domain-specific knowledge in law, medicine, has absorbed about the world during training on vast
or advanced mathematics. The historical inspiration text corpora.
for MMLU drew on the observation that human The benchmark comprises 57 tasks that span a
education spans a broad curriculum, and compelling wide array of subjects across four broad categories:
AI systems should be able to handle questions from Humanities, Social Sciences, STEM, and Other
any part of that curriculum. The developers of domains [24]. Table 3 lists these categories
MMLU sought to assemble a benchmark that would: with example subjects. Each task is designed
as a multiple-choice question-answering problem,
• Span Diverse Domains: Incorporate tasks from typically with four options per question. This format
STEM fields, social sciences, humanities, and was chosen for several reasons:
other disciplines, reflecting the breadth of human
knowledge. • Multiple-choice questions are common in
standardized tests and allow objective grading
• Cover Different Difficulty Levels: Include via accuracy.
problems ranging from elementary school
to professional exam level, ensuring that • The fixed choice format (with a 25% random
the benchmark tests basic knowledge and guess baseline) provides a clear measure
expert-level reasoning. of improvement as models exceed chance
performance.
• Remain Robust to Short-Term Saturation:
Provide a large and varied challenge such that • It enables evaluation of factual recall and
models would likely require significant advances problem-solving, as questions can be conceptual
to excel, preventing immediate saturation as with (testing knowledge) or analytical (requiring
GLUE. reasoning to eliminate distractors).
• Evaluate Multitask Generalization: Emphasize Another theoretical underpinning of MMLU is
a model’s ability to handle many tasks without its granularity and difficulty stratification. Many

ISSN: 2367-8895 84 Volume 10, 2025

 International Journal of Computers
Maikel Leon [Link]

and professional exams like the United States Medical

Table 3: Broad Categories of MMLU Tasks with Licensing Examination (USMLE) [24]. This curation
Examples strategy ensured that the content of MMLU is
Category Example Subjects (Task) realistic and representative of the types of questions
Humanities World History, Law, Philosophy a well-educated human might encounter.
Social Sciences Psychology, Economics,
Political Science Each subject in MMLU is represented by a set
STEM Mathematics (elementary to of multiple-choice questions, typically with four
college), Physics, Computer answer choices labeled (A), (B), (C), and (D). The
Science dataset is further partitioned into a small development
Other Domains Medicine (USMLE-style), set, a validation set, and a held-out test set. The
Business, Ethics development set includes a few example questions per
subject, meant to be used for few-shot prompting. The
validation set can select model hyperparameters or
subjects appear at multiple levels (for instance, evaluate prompts, while the test set is used for the final
mathematics has separate tasks for elementary, high benchmark evaluation. Notably, the test set for each
school, and college math) [24]. This design allows subject contains a substantial number of questions
analysis of a model’s progress as questions become (often in the order of 100 or more), making the
more advanced, mirroring the human learning assessment statistically reliable and reducing variance
trajectory in those subjects. Similarly, some [24].
professional domains (law, medicine) are included to
test specialized knowledge and reasoning akin to what The primary evaluation metric for MMLU is
a trained expert would possess. accuracy, the percentage of questions answered
MMLU’s emphasis on zero-shot and few-shot correctly. Because each question has four options,
evaluation ties into the theoretical concept of a naive baseline achieves 25% accuracy on average.
”few-shot generalization”. The creators explicitly MMLU results are often reported in two forms:
avoided fine-tuning models on these tasks for overall accuracy (across all questions in all tasks)
benchmark scoring; instead, models are prompted and per-category or per-task accuracy. The overall
with zero or a few exemplars from a task and then score can be either a micro-average (weighing
must answer new questions [24]. This protocol each question equally) or a macro-average across
measures how well models can generalize knowledge functions; in the original work, it was reported as a
without gradient-based learning on the target task, weighted average accuracy to aggregate performance
echoing how humans apply general knowledge to [24]. They also break down results by the four broad
unfamiliar problems. This design choice makes categories (as in Table 3) to diagnose models’ relative
MMLU a stringent test of a model’s inherent strengths and weaknesses in different knowledge
capabilities derived from pretraining rather than its domains.
ability to learn from additional supervised data on
the benchmark itself. In summary, the theoretical Models are evaluated in either a zero-shot setting,
foundation of MMLU is the notion of comprehensive, where the model is given only the question (and
transferrable language understanding. MMLU perhaps the subject’s statement), or a few-shot setting,
is intended to be a reliable proxy for a model’s where a handful of example Q&A pairs from the same
real-world knowledge, competence, and reasoning subject are provided as a prompt prefix. For example,
skills across domains by covering a broad knowledge a few-shot prompt might begin with:
spectrum and enforcing evaluation conditions Subject: High School Physics.
analogous to how humans tackle standardized tests. Q1: (question text)
A. ... B. ... C. ... D. ...
3.3 Dataset Composition and Evaluation Answer: B
The MMLU dataset comprises approximately 16,000 Q2: (question text) ...
question-answer pairs divided among the 57 tasks Answer: ...
[24]. These questions were primarily from publicly Q3: (new question)
available resources, such as practice exams and study Answer:
materials for various academic tests and professional This format tests the model’s ability to follow the
certifications. For instance, a portion of the questions pattern and answer the new question. Notably, in the
come from Graduate Record Examination (GRE) original benchmark definition, no gradient updates or
practice sets, Advanced Placement (AP) course fine-tuning on MMLU are performed; the model must
exams (high school level), undergraduate curricula, use its pre-existing knowledge.

ISSN: 2367-8895 85 Volume 10, 2025

 International Journal of Computers
Maikel Leon [Link]

3.4 Performance and Results researchers have noted that specific MMLU subjects
MMLU exposed significant gaps between remain difficult and that the benchmark itself has
contemporary models and human experts upon its limitations (e.g., some questions are ambiguously
introduction. Table 4 summarizes the performance of worded or have erroneous answers) [30]. An audit
several models on the MMLU test (few-shot setting). of MMLU questions identified errors in about 6.5%
Early transformer-based models like RoBERTa and of the questions, implying that even an ideal model
ALBERT barely improved over chance. GPT-3, might max out below 100% on this benchmark [30].
with 175 billion parameters [25], was the first model This finding suggests that further gains need careful
to substantially outperform random guessing on interpretation as models approach the 90% range.
MMLU, achieving around 44% accuracy overall.
This was a notable jump, yet it still fell far short 3.5 Applications and Impact
of human expert performance, estimated at around Although MMLU is a benchmark rather than an
90% [24]. application, it has a significant practical impact on the
development and deployment of language models:
• Model Benchmarking: MMLU is now a
Table 4: Example MMLU Accuracy Results
routinely reported metric in major language
(Few-Shot) from early evaluations [24]
model releases. It allows researchers and
Model Hum. Soc. Sci. STEM Avg. industry practitioners to compare models in
Random terms of broad knowledge and reasoning,
(25% b-l) 25.0 25.0 25.0 25.0 much like an IQ test for AI. For example,
RoBERTa academic papers and industrial reports (OpenAI,
(fine-tuned) 27.9 28.8 27.0 27.9 DeepMind, Anthropic, etc.) use MMLU to
UnifiedQA demonstrate a model’s strengths and weaknesses
(T5-based) 45.6 54.6 40.2 48.9 across subjects.
GPT-3 • Diagnostic for Weaknesses: The granularity
(175B, f-s) 40.8 50.4 36.7 43.9 of MMLU (with per-subject results) helps
Human identify domains where a model may be lacking.
(expert est.) – – – ∼90.0 If a model performs poorly in economics
or mathematics relative to other areas, this
The results in Table 4 illustrate several noteworthy can guide targeted improvements or additional
points. First, performance varies by category: training data. In this sense, MMLU informs the
for instance, GPT-3 was relatively stronger on iterative design of more robust AI systems.
humanities and social sciences questions than on
• Real-world Readiness: Success on MMLU
STEM questions, echoing the observation that models
correlates with a model’s ability to handle
tend to find calculation-heavy or formal reasoning
knowledge-intensive tasks. For instance, a
tasks (math, physics) more challenging than factual
model that scores highly on medical and law
or text-based tasks. Second, the large gap between questions in MMLU might be more reliable for
GPT-3 and the human expert level underscored how assisting in those domains (though it would still
far even the best model in 2020 was from robust require careful validation). In effect, MMLU
multidisciplinary understanding. is a proxy for how well a model has absorbed
In the years since MMLU’s release, it has become the knowledge a human professional or student
a standard evaluation for new large language models. would need, which is relevant when considering
Progress has been remarkable: by 2022–2023, AI for educational tools, expert systems, or
models like DeepMind’s Chinchilla (70B) and decision support.
Google’s PaLM (540B) reached scores in the 60–70%
range on MMLU [26, 27], and by 2023–2024, • Research on General Intelligence: As a
cutting-edge models such as GPT-4 reportedly scored comprehensive test, MMLU feeds into
around 86% in a zero-shot setting on MMLU discussions about artificial general intelligence.
[28], and nearly 90% with advanced prompting or An AI system’s performance on MMLU
fine-tuning techniques [29]. This approaches the provides a single-number summary of its
estimated human expert performance, a milestone general academic competency. This has been
that just a few years prior seemed distant. Such cited in debates about whether models truly
improvements reflect the increasing scale of models understand content or merely recall it, and how
and enhancements in training methods that better far current models are from human-like breadth
capture knowledge and reasoning. At the same time, of cognition.

ISSN: 2367-8895 86 Volume 10, 2025

 ll
OPEN ACCESS

Article
Worldwide AI ethics: A review
of 200 guidelines and recommendations
for AI governance
Nicholas Kluge Correˆ a,1,2,7 Camila Galva˜ o,3 James William Santos,1,* Carolina Del Pino,4 Edson Pontes Pinto,3
Camila Barbosa,1 Diogo Massmann,1 Rodrigo Mambrini,3 Luiza Galva˜ o,5 Edmund Terem,6 and Nythamar de Oliveira1
1Graduate Program in Philosophy, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
2
Graduate Program in Philosophy, University of Bonn, Bonn, North Rhine-Westphalia, Germany
3Graduate Program in Law, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
4Psychology Undergraduate Program, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
5
Law Undergraduate Program, Pontifical Catholic University of Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
6Graduate Program in Philosophy, University of Johannesburg, Porto Alegre, Johannesburg, South Africa
7Lead contact

*Correspondence: [Link]@[Link]
[Link]

THE BIGGER PICTURE After the AI winter in the late 80s, AI research has experienced remarkable growth.
Currently, a lot of work is taking place to define the values and ideas that should guide AI advances. A key
challenge, however, lies in establishing a consensus on these values, given the diverse perspectives of
various stakeholders worldwide and the abstraction of normative discourse. Researchers and policy
makers need better tools to catalog and compare AI governance documents from around the world and
to identify points of divergence and commonality.

SUMMARY

The utilization of artificial intelligence (AI) applications has experienced tremendous growth in recent years,
bringing forth numerous benefits and conveniences. However, this expansion has also provoked ethical con-
cerns, such as privacy breaches, algorithmic discrimination, security and reliability issues, transparency, and
other unintended consequences. To determine whether a global consensus exists regarding the ethical prin-
ciples that should govern AI applications and to contribute to the formation of future regulations, this paper
conducts a meta-analysis of 200 governance policies and ethical guidelines for AI usage published by public
bodies, academic institutions, private companies, and civil society organizations worldwide. We identified at
least 17 resonating principles prevalent in the policies and guidelines of our dataset, released as an open
source database and tool. We present the limitations of performing a global-scale analysis study paired
with a critical analysis of our findings, presenting areas of consensus that should be incorporated into future
regulatory efforts.

INTRODUCTION puter science, the most frequently submitted sub-categories

for publications are "computer vision and pattern recognition,"
Since the last period of reduced interest in artificial intelligence "machine learning," and "computation and language," 1 i.e.,
(AI), the "AI winter" from 1987 to 1993, the field of AI research areas where machine learning (a sub-field of AI research) has es-
and industry has witnessed significant growth. This growth en- tablished itself as the reigning paradigm.
compasses various aspects, including the development of new Moreover, investment in AI-related companies and startups
technologies, increased investment, greater media attention, has reached unprecedented levels, with governments and ven-
and expanded capabilities of autonomous systems. A study ture capital firms investing over $90 billion (USD) in the United
analyzing the submission history on ArXiv from 2009 to 2021 re- States alone in 2021, accompanied by a surge in the registration
veals that computer science-related articles have become the of AI-related patents.2 While these money-field advancements
most prevalent type of material submitted, increasing 10-fold have brought numerous benefits, they also introduce risks and
starting in 2018. Furthermore, within the broad scope of com- side effects that have promoted several ethical concerns, like

Patterns 4, 100857, October 13, 2023 ª 2023 The Author(s). 1

This is an open access article under the CC BY license ([Link]
 ll
OPEN ACCESS Article
risks to user privacy, the potential for increased surveillance, the stitutions (e.g., AI Now Institute17), and professional associations
environmental cost of the industry, and the amplification of prej- (e.g., IEEE18), among other types of institutions, representing a
udices and discrimination on a large scale, which can dispropor- multi-stakeholder sample that for years was the most extensive
tionately harm vulnerable groups. Consequently, the expansion collection of AI guidelines analyzed systematically.
of the AI industry has given rise to what we refer to as the "AI One of the main findings in Jobin et al.’s3 work was the detec-
ethics boom," i.e., a period marked by an unprecedented de- tion of the most common ethical principles in the discourse of the
mand for regulation and normative guidance in this field. evaluated documents, those being transparency, justice/equity,
One of the central questions surrounding this boom is the non-maleficence, accountability, privacy, beneficence, freedom
determination of what ethical premises should guide the devel- and autonomy, trust, dignity, sustainability, and solidarity. Of
opment of AI technologies. And to answer this question, a these 11 ethical principles cited, five were the most recurrent:
plethora of principles and guidelines have been proposed by transparency (86%), justice (81%), non-maleficence (71%), re-
many stakeholders. However, the consensus and divergences sponsibility (71%), and privacy (56%). Furthermore, Jobin
in these varied discourses have yet to be extensively accessed. et al.’s work is careful not to impose any normative guidelines
For instance, do Silicon Valley-based companies follow the for the effectiveness of the mentioned principles. It attempts to
same precautions as major Chinese technology firms? Are these raise the issue and map the global picture in a pioneer work of
concerns relevant to end-users in countries with diverse cultural descriptive ethics. However, their limited sample, where, for
and social norms? Establishing a consensus to support the example, no Latin American countries are mentioned, makes
global regulations currently under discussion is of paramount the true global representativeness of the results questionable,
importance in both a practical and theoretical sense. a limitation recognized by Jobin et al. as one of their blind spots.
To address these questions, we draw inspiration from previ- Thilo Hagendorff19 conducted another study that presented a
ous works by meta-analysts and meticulously survey a wide similar type of analysis. In his research, Hagendorff focused on
array of available ethical guidelines related to AI development. a smaller sample of 21 documents. He excluded those older
These sources include governance policies of private com- than 5 years and that only addressed a national context unrelated
panies, academic institutions, and governmental and non- to AI, such as data science and robotics. Furthermore, Hagendorff
governmental organizations, as well as ethical guidelines for AI did not consider corporate policies but deliberately selected doc-
usage. By analyzing 200 documents in five different languages, uments deemed relevant in the international discourse (IEEE,
we gathered information on what ethical principles are the Google, Microsoft, and IBM) based on his evaluation criteria.
most popular, how they are described, where they come from, Even working with a smaller sample, Hagendorff’s findings
their intrinsic characteristics, and much else. Our primary goal corroborate with those of Jobin et al.,3 where the most mentioned
was to identify the most advocated ethical principles, to examine principles found were accountability (77%), privacy (77%), justice
their global distribution, and to assess if there is a consistent un- (77%), and transparency (68%). Hagendorff (like Jobin et al.) also
derstanding of these principles. Ultimately, this analysis aims to mentions the underrepresentation of institutions in South America,
determine whether a consensus exists regarding the normative Africa, and the Middle East as a clear bias of his sample.
discourse presented in ethical guidelines surrounding AI
Hagendorff is also more critical in his analysis, presenting
development.
concluding points such as the following:

State of the art d The lack of attention given to questions related to labor
One of the first studies to promote a meta-analysis of published rights, technological unemployment, the militarization of
AI ethical guidelines was that of Jobin et al.3 In this study, these AI, the spread of disinformation, and the misuse/dual-use
authors sought to investigate whether a global agreement on of AI technology
emerging questions related to AI ethics and governance would d The lack of gender diversity in the tech industry and AI
arise. The research identified 84 documents containing ethical ethics
guidelines for intelligent autonomous systems using the d The short, brief, and minimalist views some documents
Preferred Reporting Items for Systematic Reviews and Meta- give to normative principles
Analyses framework (originally developed for analysis in the d The lack of technical implementations for how to imple-
healthcare sector).4 At the time, some of them were one of the ment the defended principles in AI development
most cited guidelines in the literature, like the Organization for d The lack of discussion on long-term risks (e.g., artificial
Economic Co-operation and Development (OECD) Recommen- general intelligence safety, existential risks)
dation of the Council on Artificial Intelligence, 5 the High-Level
Expert Group on AI Ethics Guidelines for Trustworthy AI,6 the While the works of Hagendorf and Jobin et al. are valuable
University of Montre´ al Declaration for responsible development contributions to AI ethics, we question the criteria for filtering
of artificial intelligence,7 the Villani Mission’s French National documents used by these works. We argue that if we want to
Strategy for AI,8 among many others. investigate the consensus regarding the normative dispositions
Jobin et al.’s3 sample also contained documents from govern- of different countries and organizations regarding AI, we should
mental organizations (e.g., Australian Government Department not use popularity-based filtering. In other words, a descriptive
of Industry Innovation and Science 9), private companies (e.g., ethics evaluation should take as many viewpoints as possible if
SAP,10 Telefonica,11 IBM12), non-governmental organizations it aims to make a solid description of the "global landscape."
(e.g., Future Advocacy,13 AI4People14), non-profit organizations As a last mention, we would like to cite the work done by Fjeld
(e.g., Internet Society,15 Future of Life Institute16), academic in- et al.,20 one of the first to present documents from Latin America.

2 Patterns 4, 100857, October 13, 2023

 ll
Article OPEN ACCESS

In their study, Fjeld et al. worked with 36 samples produced by a gendorff and Fjeld et al.20 was not "user-friendly" or clear,
variety of institution types, but as with Hagendorff, 19 Fjeld et al. something that we tried to overcome in our work.
also excluded data science, robotics, and other AI-related (4) Released with an open source dataset, making our work
fields/applications. According to these authors, eight principles reproducible and extendable.
were the most commonly cited in their sample: fairness/non-
discrimination (present in 100% of the analyzed documents), pri- The focus of this study is guidelines related to the ethical use of
vacy (97%), accountability (97%), transparency/explainability AI technologies. We refer to "guidelines" as documents concep-
(94%), safety/security (81%), professional responsibility (78%), tualized as recommendations, policy frameworks, legal land-
human control of technology (69%), and promotion of human marks, codes of conduct, practical guides, tools, or AI principles
values (69%). for the use and development of this type of technology. The reso-
Like in the study of Jobin et al.,3 Fjeld et al.20 cite the variability in nating foundation for most of these documents is the presence of
how such principles are defined as an important element to be ad- a form of ‘‘principlism,’’24 i.e., the use of ethical principles to sup-
dressed. For example, in the 2018 version of the Chinese Artificial port normative claims.
Intelligence Standardization White Paper,21 the authors mention Now, deconstructing the expression "AI technologies," with
that AI can serve to obtain more information from the population, "AI," our scope of interest encompasses areas that inhabit the
even beyond the data that has been consented to (i.e., violation of multidisciplinary umbrella that is artificial intelligence research,
informed consent would not undermine the principle of privacy), such as statistical learning, data science, machine learning (ML),
while the Indian National Strategy for Artificial Intelligence Discus- logic programming/symbolic AI, optimization theory, robotics,
sion Paper (National Institution for Transforming India)22 argues software development/engineering, etc. And with the term "tech-
that its population must become massively aware so that they nologies," we refer to specific tools/techniques, applications, and
can effectively consent to the collection of personal information. services. Thus, the term refers to technologies for automating
We consider the delineation of diverging principles across docu- decision processes and mimicking intelligent/expert behavior.
ments to be one of the standout strengths of Fjeld et al.’s work, Unlike previous works,19,20 we included areas disregarded previ-
which warrants replication on a broader scale. ously (software development, data science, robotics).
There is more meta-analytical work that has been done in AI
ethics that we will not cite in depth. For a complete review of Sources
meta-analytical research on normative AI documents, we We used as sources for our sample two public repositories, the
recommend the work done by Schiff et al., 23 which cites many ‘‘AI Ethics Guidelines Global Inventory,’’ from AlgorithmWatch,
other important works. and the ‘‘Linking Artificial Intelligence Principles’’ (LAIP) guide-
Now, shifting the perspective from past works to our own, we lines. The AlgorithmWatch repository contained 167 documents,
argue that many of the mentioned analyses, following Jobin while the LAIP repository contained 90.
et al.’s3 work, suffer from a small sample size. While works like Initially, we checked for duplicate samples between both re-
the ones done by Hagendorff19 and Fjeld et al.20 show more positories. After disregarding them, we also scavenged for
diversified categories of typologies or sample sizes, their limited more documents through web search engines and web
sample may hinder the generalization of their results. Meanwhile, scraping, utilizing search strings such as "Artificial Intelligence
Jobin et al. do not present some features (e.g., an extensive Principles," "Artificial Intelligence Guidelines," "Artificial Intelli-
exemplification of how principles diverge) presented in other gence Framework," "Artificial Intelligence Ethics," "Robotics
works that used a smaller sample. Also, we would like to point Ethics," "Data Ethics," "Software Ethics," and "Artificial Intelli-
out that none of these studies released their dataset in a form gence Code of Conduct," among other related search strings.
that would allow the replication of their findings, which makes We limited our search to samples written/translated in one of
all of these studies (if one would not be willing to redo all the the five languages our team could cope with: English, Portu-
work) irreproducible. guese, French, German, and Spanish.
To cover these shortcomings, we propose an updated review We were able to collect in this manner 200 documents. Thus,
of AI guidelines and related literature: Worldwide AI Ethics. after defining our pool of samples, we initiated the data collection
stage. We divided this stage into two phases.
Methodology
From the gaps pointed out in the previous meta-analyses, in this First phase
study, we sought to present the following to the AI community. In phase one, members of our team received different quotas of
documents. Each team member was responsible for reading,
(1) A quantitatively larger and more diverse sample size, as translating when needed, and hand-coding pre-established fea-
done by Jobin et al.3 Our sample possesses 200 docu- tures. The first features looked for were the following:
ments originating from 37 countries, spread over six con-
tinents, in five different languages. d Institution responsible for producing the document
(2) Combined with a more granular typology of document d Country/world region of the institution
types, as done by Hagendorff.19 This typology allowed d Type of institution (e.g., academic, non-profit, govern-
an analysis beyond the mere quantitative regarding the mental, etc.)
content of these documents. d Year of publication
(3) Presented in an insightful data visualization framework. d Principles (as done by Fjeld et al.,20 we broke principles
We believe the data presentation done by authors like Ha- into themes of resonating discourse)

Patterns 4, 100857, October 13, 2023 3

 ll
OPEN ACCESS Article
d Principles description (i.e., the words used in a document d Self-regulation/voluntary self-commitment: this category
to define/support a given principle) is designed to encompass documents made by private
d Gender distribution among authors (inferred through a first organizations and other bodies that defend a form of
name automated analysis) self-regulation governed by the AI industry itself. It also en-
d Size of the document (i.e., word count) compasses voluntary self-commitment made by indepen-
dent organizations (NGOs, professional associations, etc.).
For the gender inference part, after removing documents with d Recommendation: this category is designed to encom-
unspecified authors, we performed a name-based gender anal-
pass documents that only suggest possible forms of
ysis. Given the variety/diversity that names can possess, it was governance and ethical principles that should guide orga-
necessary to use automation to infer gender encodes (male/fe- nizations seeking to use, develop, or regulate AI tech-
male). To make an accurate inference, we also extracted (in addi- nologies.
tion to each author’s name) the most likely nationality associated
with each name. For this, we used (in addition to the country/origin We defined these categories as mutually exclusive (the pres-
of each document) an API service that predicts the most likely na- ence of a feature excludes the other). The third type of document
tionality associated with a given name. Finally, we used another classification pertains to the normative strength of the proposed
API service to infer gender based on first name plus nationality. regulation mechanism. In this regard, we established two distinct
You can find the code for our implementation in this repository: categories, drawing upon the definitions provided by the ‘‘Inno-
[Link] vative and Trustworthy AI.’’25
Regarding how we defined principles, in the first phase, we es-
tablished a list of principles so our team could focus their search. d Legally non-binding guidelines: these documents propose
We based these on past works mentioned in the "related works" an approach that intertwines AI principles with recommen-
section: accessibility, accountability, auditability, beneficence/ ded practices for companies and other entities (i.e., soft
non-maleficence, dignity, diversity, freedom/autonomy, hu- law solutions).
man-centeredness, inclusion, intellectual property, justice/eq- d Legally binding horizontal regulations: these documents
uity, open source/fair competition, privacy, reliability, solidarity, propose an approach that focuses on regulating specific
sustainability, and transparency/explainability. uses of AI through legally binding horizontal regulations,
We also used this phase to determine categories/types as- such as mandatory requirements and prohibitions.
signed to each document in the second phase. These types
were determined by the following: We defined them as mutually inclusive. The final type relates to
the perceived impact scope that motivates the evaluated docu-
(1) The nature/content of the document ment. With impact scope, we mean the dangers and negative
(2) The type of regulation that the document proposes prospects regarding the use of AI that inspired the type of
(3) The normative strength of this regulation normative propositions used. For this, three final categories
(4) The impact scope that motivates the document were defined and also posed as mutually exclusive.

The first type relates to the nature/content of the document. d Short-termism: we designed this category to encompass
documents in which the scope of impact and preoccupa-
d Descriptive: descriptive documents take the effort of pre-
tion focus mainly on short-term problems, i.e., problems
senting factual definitions related to AI technologies. These we are facing with current AI technologies (e.g., algorithmic
definitions serve to contextualize "what we mean" when
discrimination, algorithmic opacity, privacy, legal account-
we talk about AI.
ability).
d Normative: normative documents present norms, ethical
d Long-termism: we designed this category to encompass
principles, recommendations, and imperative affirmations
documents in which the scope of impact and preoccupa-
about what such technologies should be used/devel-
tion focus mainly on long-term problems, i.e., problems
oped for.
we may come to face with future AI systems. Since such
d Practical: practical documents present development tools
technologies are not yet a reality, we can classify these
to implement ethical principles and norms, be they qualita- risks as hypothetical or, at best, uncertain (e.g., sentient
tive (e.g., self-assessment surveys) or quantitative (e.g.,
AI, misaligned AGI, super-intelligent AI, AI-related existen-
debiasing algorithms for ML models).
tial risks).
We defined these first three categories as mutually inclusive d Short-termism and long-termism: we designed this cate-
(documents may have all these features combined). The second gory to encompass documents in which the scope of
type relates to the form of regulation that the document impact is both short and long-term, i.e., they present a
proposes. "mid-term" scope of preoccupation. These documents
address issues related to the short-termism category while
d Government regulation: this category is designed to also pointing out the mid/long-term impacts of our current
encompass documents made by governmental institutions AI adoption (e.g., AI interfering in democratic processes,
to regulate the use and development of AI, strictly (legally autonomous weapons, existential risks, environmental
binding horizontal regulations) or softly (legally non-binding sustainability, labor displacement, and the need for updat-
guidelines). ing our educational systems).

4 Patterns 4, 100857, October 13, 2023

 ll
OPEN ACCESS Article

autonomy of human decision-making must be preserved parency of an organization" or "the transparency of an al-
during human-AI interactions, whether that choice is indi- gorithm." This set of principles is also related to the idea
vidual or the freedom to choose together, such as the invi- that such information should be understandable to nonex-
olability of democratic rights and values, also being linked perts and, when necessary, subject to be audited.
to technological self-sufficiency of nations/states. d Truthfulness: this principle upholds the idea that AI tech-
d Human formation/education: such principles defend that nologies must provide truthful information. It is also related
human formation and education must be prioritized in our to the idea that people should not be deceived when inter-
technological advances. AI technologies require consider- acting with AI systems.
able expertise to be produced and operated, and such
knowledge should be accessible to all. These 17 principles contemplate all of the normative discourse
d Human-centeredness/alignment: such principles advo- we could interpret. At the end of this second phase, all docu-
ments received 13 features: origin country, world region, institu-
cate that AI systems should be centered on and aligned
tion, institution type, year of publication, principles, principles
with human values. This principle is also used as a
"catch-all" category, many times being defined as a collec- definition, gender distribution, size, type I (nature/content),
tion of "principles that are valued by humans" (e.g., type II (form of regulation), type III (normative strength), type IV
freedom, privacy, non-discrimination, etc.). (impact scope), plus identifiers and attachments like document
d Intellectual property: this principle seeks to ground the title, abstract, document URL, and related documents.
property rights over AI products and their generated
outputs. Our tool
d Justice/equity/fairness/non-discrimination: this set of prin- We used all information obtained during the second phase
ciples upholds the idea of non-discrimination and bias miti- to create the dataset that feeds our visualization tool: an interac-
gation (discriminatory algorithmic biases AI systems can tive dashboard. We created this dashboard using the Power
be subject to). It defends that, regardless of the different BItool ([Link]
sensitive attributes that may characterize an individual, [Link]). We also developed a secondary dashboard
algorithmic treatment should happen "fairly." (open source) using the Dash library ([Link]
d Labor rights: labor rights are legal and human rights related [Link]/worldwide_AI-ethics/), an open source framework
to the labor relations between workers and employers. In for building data visualization interfaces.
AI ethics, this principle emphasizes that workers’ rights The main distinction between our tool and Hagendorff’s
should be preserved regardless of whether labor relations table19 and Fjeld et al.’s graphs20 lies in its interactivity and the
are being mediated/augmented by AI technologies. flexibility to combine various filters without being confined to
d Cooperation/fair competition/open source: this set of prin- preconfigured orderings. This feature allows researchers to uti-
ciples advocates different means by which joint actions lize the tool to examine and question specific characteristics pre-
can be established and cultivated between AI stakeholders sent in their regions, to identify trends and behaviors, or to
to achieve common goals. It also relates to the free and explore categories relevant to their research focus. It is worth
open exchange of valuable AI assets (e.g., data, knowl- noting that we were among the first to openly release our data-
edge, patent rights, human resources). set, making our work accessible and reproducible.
d Privacy: the idea of privacy can be defined as the individ- Another distinguishing feature of our tool is its ability to
ual’s right to "expose oneself voluntarily, and to the extent condense large amounts of information into a single visualization
desired, to the world." This principle is also related to data panel. Our choice for such a way of presenting our data was to
protection related-concepts such as data minimization, make it easier to interpret how certain features interact with
anonymity, informed consent, and others. others. While previous works demonstrate the statistical distri-
d Reliability/safety/security/trustworthiness: this set of prin- bution of certain features, ours allows the user to see how these
ciples upholds the idea that AI technologies should be reli- features vary and are interconnected.
able, in the sense that their use can be truly attested as safe
and robust, promoting user trust and better acceptance of Limitations
AI technologies. As in past works, this analysis also suffers from a small sample.
d Sustainability: this principle can be interpreted as a mani- Our work represents a mere fraction of what our true global
festation of "intergenerational justice," wherein the welfare landscape on this matter is. Some of the main limitations we
of future generations must be considered in AI develop- encountered during our work are the following.
ment. In AI ethics, sustainability pertains to the notion
that the advancement of AI technologies should be ap- (1) The limited scope of languages we were able to interpret
proached with an understanding of their enduring conse- represents a language bias, potentially excluding relevant
quences, encompassing factors such as environmental perspectives.
impact and the preservation and well-being of non-human (2) Publication bias is also a concern, as the focus on pub-
life. lished guidelines may overlook valuable insights from
d Transparency/explainability/auditability: this set of princi- ongoing discussions in other forms of media.
ples supports the idea that the use and development of (3) The "guideline" scope excludes the academic work being
AI technologies should be transparent for all interested done worldwide (i.e., we did not consider academic pa-
stakeholders. Transparency can be related to "the trans- pers on AI ethics).

6 Patterns 4, 100857, October 13, 2023

 ll
OPEN ACCESS Article

personal data (without consent) allowed for personal profiling choose to define AI as only "systems that can learn," you will
and targeted political advertising during elections. 42,43 We can leave outside your scope of regulation an entire family of systems
also mention relevant works that helped cement the AI ethics that do not learn (rule-based systems) but can still act "intelli-
field as a popular area of research, like the book Weapons of gently" and autonomously.
Math Destruction.44 Meanwhile, as already stated by Fjeld et al., 20 there is a gap
All these events and many others may have helped bring this between established principles and their actual application. In
burst of interest to the field. Perhaps some of these events could our sample, most of the documents only prescribe normative
come to explain the swings of attention on AI ethics. And of claims without the means to achieve them, while the effective-
course, this increase in popularity may also be related to the ness of more practical methodologies, in the majority of cases,
increased funding that AI research (where AI ethics remains as remains extra empirical. With this, we see a field with a significant
a sub-field) received in the last decade.2 lack of practical implementations46 that could support its norma-
tive claims.
Typological categories This fact may become more alarming when we look at the dis-
Regarding the previously defined typological categories, when tribution of government documents that opt for "soft" forms of
looking at the document’s nature/content, we found that the ma- regulation (91.6%). The critique that "ethical principles are not
jority of our sample is from the normative type (96%), which a enough to govern the AI industry" is not a new one.3,19,47–49 How-
third of the time also presents descriptive contents (55.5%) ever, perhaps those critiques have not yet permeated the main-
and, more rarely, practical implementations (2%). stream community, which, by our analysis, is still largely based
When we look at the form of regulation proposed by the on principles detached from observable metrics or practical
documents of our sample, more than half (56%) are only recom- implementations.
mendations to different AI stakeholders, while 24% possess self- However, even if most countries in our sample seem to opt for
regulatory/voluntary self-commitment style guidelines, and only legally non-binding forms of regulation, there seems to be a
20% propose a form of regulation administered by a given state/ growing adoption/proposition of stricter solutions. The idea
country. that "ethics" and "compliance" are separate domains seems to
This lack of convergence to a more "government-based" form get ever-growing acknowledgment by countries such as Can-
of regulation reflects in the normative strength of these docu- ada, Germany, and the United Kingdom (which comprise
ments, where the vast majority (98%) only serve as "soft laws," 66.6% of our total "legally binding" sample), while according to
i.e., guidelines that do not entail any form of a legal obligation, Zhang et al., the legislative records on AI-related bills grew
while only 4.5% propose stricter regulation. Since only gov- from just one in 2016 to 18 in 2021, with Spain, the United
ernmental institutions can create legally binding norms (other in- Kingdom, and the United States being the top three "proto-AI-
stitutions lack this power), and they produced only 24% of our legislators" from 2021, showing that in fact, we may be passing
sample, some may argue that this imbalance lies in this fact. into a transitioning phase where these principles may soon be
However, by filtering only the documents produced by govern- transformed into actual laws.
mental institutions, the disproportion does not go away, with We would also like to point out the seemingly low attention
only 18.7% of samples proposing legally binding forms of regu- given to the long-term impacts of AI (1.5%). Even though there
lation. The countries on the front of this still weak trend are Can- is a considerable amount of work produced on the matter, 50–55
ada, Germany, and the United Kingdom, with Australia, Norway, many times the terms "safety," "alignment," or "human-level
and the United States coming right behind. AI" are generically dismissed as not serious or as Stuart Rus-
Our last typology group is impact scope. Looking at the totality sell56 would say, "myths and moonshine." Possible explanations
of our sample size, we see that short-term (47%) and "mid-term" for this fact could be the following: (1) the AI community does not
(i.e., short-term and long-term = 52%) prevail over more long- find these problems real; (2) the AI community does not find
term preoccupations (2%). When we filter our sample by impact these problems urgent; (3) the AI community thinks we have
scope and institution type, it seems to us that private corpora- more urgent problems at hand; or even (4) that the AI community
tions think more about the short-term (33%), governmental insti- does not know about such issues. Regardless of their urgency,
tutions about the short/long-term (28%), and academic (66%) we argue that the current lack of attention given to safety-related
and non-profit organizations (33%) with the long-term impacts topics in the field is alarming. For example, if we look at the
of AI technologies. distribution of papers submitted in the NeurIPS 2021, approxi-
mately 2% were safety related (e.g., AI safety, ML fairness, pri-
Definitions, lack of tools, the legislative push, and vacy, interpretability).
uncertain risks
In regard to the nature/content of our samples, we see that only Principle distribution
55.5% of documents (111) seek to define what is the object of Examining the distribution of principles among our total sample
their discourse, i.e., "we are talking about autonomous intelligent size, we arrive at the following results: the top five principles
systems, and this is what we understand as an autonomous advocated in the documents of our sample are similar to the re-
intelligent system." This is a curious phenomenon, more so if sults shown by Jobin et al.3 and Hagendorff,19 with the addition
we acknowledge that there is no consensual definition of what of reliability/safety/security/trustworthiness (78%), which also
"artificial intelligence" is and what it is not.45 There are many in- was top five in Fjeld et al.’s20 meta-analysis (80%) (Figure 5).
terpretations and contesting definitions, which may prove to be a Looking at principle distribution filtered by continent, the top
challenge for regulating organizations. For example, if you five principles remain the same in both North America and

10 Patterns 4, 100857, October 13, 2023

 Ethics of AI: A Systematic Literature Review of Principles and
Challenges
Arif Ali Khan1∗, Sher Badshah2, Peng Liang3, Muhammad Waseem3, Bilal Khan4, Aakash Ahmad5,
Mahdi Fahmideh6, Mahmood Niazi7, Muhammad Azeem Akbar8
1
M3S Empirical Software Engineering Research Unit, University of Oulu, Oulu, Finland
2
Faculty of Computer Science, Dalhousie University, Halifax, Canada
3
School of Computer Science, Wuhan University, Wuhan, China
4
Department of Computer Science, University of Loralai, Balochistan, Pakista
5
College of Computer Science and Engineering, University of Ha’il, Saudi Arabia
6
School of Business at University of Southern Queensland, Queensland, Australia
7
Department of Information and Computer Science, King Fahd University of Petroleum and Minerals, Saudi Arabia
8
Software Engineering Department, Lappeenranta-Lahti University of Technology, 53851 Lappeenranta, Finland
[Link]@[Link], sh545346@[Link], liangp@[Link], [Link]@[Link], bilal_nasar@[Link],
[Link]@[Link], [Link]@[Link], mkniazi@[Link], [Link]@[Link]
ABSTRACT ACM Reference Format:
Ethics in AI becomes a global topic of interest for both policymakers Arif Ali Khan1∗, Sher Badshah2, Peng Liang3, Muhammad Waseem3, Bilal
Khan4, Aakash Ahmad5, Mahdi Fahmideh6, Mahmood Niazi7, Muhammad
and academic researchers. In the last few years, various research
Azeem Akbar8. 2022. Ethics of AI: A Systematic Literature Review of Princi-
organizations, lawyers, think tankers, and regulatory bodies get in- ples and Challenges. In Proceedings of ACM Conference (EASE). ACM, New
volved in developing AI ethics guidelines and principles. However, York, NY, USA, 10 pages. [Link]
there is still debate about the implications of these principles. We
conducted a systematic literature review (SLR) study to investigate 1 INTRODUCTION
the agreement on the significance of AI principles and identify the
Artificial intelligence (AI) technologies are considered important
challenging factors that could negatively impact the adoption of AI
across a vast array of industries including health, manufacturing,
ethics principles. The results reveal that the global convergence set
banking and retail [14]. However, the promises of AI systems like
consists of 22 ethical principles and 15 challenges. Transparency,
improving productivity, reducing costs, and safety has now been
privacy, accountability and fairness are identified as the most com-
considered with worries, that these complex systems might bring
mon AI ethics principles. Similarly, lack of ethical knowledge and
more ethical harm than economical good [14].
vague principles are reported as the significant challenges for con-
Artificial intelligence (AI) and autonomous systems have a signif-
sidering ethics in AI. The findings of this study are the preliminary
icant effect on the development of humanity [12]. The autonomous
inputs for proposing a maturity model that assesses the ethical
decision-making nature of these systems raises fundamental ques-
capabilities of AI systems and provides best practices for further
tions i.e., what are the potential risks involved in those systems,
improvements.
how these systems should perform, how to control such systems
and what to do with AI-based systems? [12]. Autonomous systems
CCS CONCEPTS are beyond the concepts of automation by characterising them with
• Software and its engineering → Human-centered comput- decision-making capabilities. The development of autonomous sys-
ing; • Computing methodologies → Artificial intelligence; • tem components, such as intelligent awareness and self-decision
Philosophical/theoretical foundations of artificial intelli- making is based on AI concepts.
gence; • Social and professional topics → Empirical studies; There is a political and ethical discussion to develop policies for
different technologies including nuclear power, manufacturing etc.
KEYWORDS to control the ethical damage they could bring. The same ethical
AI Ethics, Machine Ethics, Principles, Challenges, Systematic Liter- potential harm also exits in AI systems and more specifically they
ature Review might end human control [12].The real-world failure and misuse
incidents of AI systems bring the demand and discussion for AI
ethics [19]. The ethical studies of AI technologies revealed that AI
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed and autonomous systems should not be only considered a techno-
for profit or commercial advantage and that copies bear this notice and the full citation logical effort. There is a broad discussion that the design and use
on the first page. Copyrights for components of this work owned by others than ACM of AI-based systems are culturally and ethically embedded [10].
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a Developing AI-based systems not only need technical efforts but
fee. Request permissions from permissions@[Link]. also include economic, political, societal, intellectual and legal as-
EASE, 13-15 June 2022, Gothenburg, Sweden pects [10]. These systems significantly impact the cultural norms
© 2022 Association for Computing Machinery.
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00 and values of the people [10]. The AI industry and specifically the
[Link] practitioners should have a deep understanding of ethics in this
EASE, 13-15 June 2022, Gothenburg, Sweden Khan et al.

domain. Recently, AI ethics get press coverage and public voices, accountability and awareness of misuse. Organizations such as ISO
which supports significant related research [19]. However, the topic and IEC also embark on developing standard for AI [5]. ISO/IEC JTC
is still not sufficiently investigated both academically and in the 1/SC 42 [5] is a joint ISO/IEC international standard committee that
real-world environment [10]. There are very few academic stud- focus on the entire AI ecosystem including ethical and social con-
ies conducted on this topic, but it is still largely unknown to AI cerns, standardization, AI governance, AI computational approach
practitioners. The Ethically Aligned Design (EAD) [7] guidelines and trustworthiness [5]. The effort of different organizations to
of IEEE mentioned that ethics in AI is still far from being mature shape AI ethics not only determine the need of guidelines, tools,
in an industrial setting [18]. The limited or no knowledge of ethics techniques, but also the interest of these organizations to manage
for the AI industry develop the gap, which indicates the need for ethics in a way that meet their respective priorities.
further academic and in practice research. However, recently published studies reported that the existing
The aim of this study is to conduct a systematic literature review guidelines developed for ethics of AI are not effective and adopted
(SLR) and explore the available literature to identify the AI ethics in practice [20]. It is evident from the empirical study conducted
principles. Moreover, the SLR study uncover the key challenging by McNamara et al. [11] to test the influence of the ACM code of
factors that are the demotivators for considering the ethics of AI. ethics in the decision-making process of software development. The
The following research questions are developed to achieve the given results of the study revealed that the ACM code of ethics have no
core objectives: impact in making ethical decisions. The lack of effective techniques
• RQ1: What are the key principles of AI ethics? makes it challenging to successfully scale the available guidelines
into practice [20]. Vakkuri et al. [20] used the accountability, respon-
• RQ2: What are the challenges of adopting ethics in AI?
sibility, and transparency (ART) framework [6] and developed the
The remaining content of paper is structured as follow: Section 2 conceptual model to explore ethical consideration in the AI environ-
presents the background of the study and the research methodology ment. The conceptual model is empirically validated by conducting
is reported in Section 3. The SLR data are provided in Section multiple case studies. The empirical results are concluded by high-
4 and results and analysis are discussed in Section 5. The study lighting that AI ethics principles are still not in practice; however,
implications are discussed in Section 6. Finally, Section 7 provides some common concepts are considered such as documentation.
an overview of threats to the validity of the study and Section 8 Moreover, the study findings revealed that practitioners consider
conclude the findings with future directions. the social impact of AI systems [20].
There are no tools, methods or frameworks that fill the gap
2 BACKGROUND between the AI principles and their implementation in practice.
The implementation of AI or machine intelligence concepts brings Further studies in this area should conduct that explicitly discuss
a technological revolution that change both science and society. the AI ethics principles, challenges and provide evaluation stan-
Human to machine power transformation sparked important soci- dards/models that guide AI industry to consider ethics in practice.
etal debate about the principles and policies that guide the use and
deployment of AI systems [8]. Various organizations have devel- 3 RESEARCH METHOD
oped ad hoc committees to draft the policy documents for AI ethics. Systematic literature review (SLR) approach is used to explore the
These organizations reportedly developed AI policies and guide available primary studies. SLR is a widely adopted literature survey
documents [8]. In 2018, technology corporates such as SAP and method in evidence-based software engineering domain. SLR is “a
Google publicly introduced guidelines and policies for AI-based means of evaluating and interpreting all available research rele-
systems [8]. Similarly, Amnesty International, the Association of vant to a particular research question, topic area, or phenomenon
Computing Machinery (ACM) and Access Now comes up with prin- of interest” [9]. The Kitchenham and Charters [9] SLR guidelines
ciples and recommendations for AI technologies. The Trustworthy are used to conduct this study and systematically address the re-
AI European Commission’s guidelines were developed with the aim search questions. The SLR process plan is provided in Figure 1 and
to promote lawful, ethically sound and robust AI systems [15]. The thoroughly discussed in the following sub-sections.
report “Preparing for the Future of Artificial Intelligence” prepared
by the Obama administration’s presents a thorough survey that 3.1 Research questions (RQs)
focuses on the current AI research, its applications and impact on
society [3]. The report further presents recommendations for future Research questions development in SLR studies is the most signifi-
AI related actions. The “Beijing AI Principles” guidelines [13] pro- cant phase [9]. Developing research questions require deep under-
posed various principles in the domain of AI research, development, standing of the research area in general and the research problem in
use and governance. These principles present a framework that specific. We primarily studied relevant articles [19][10][18][7][8][15][3]
focus on AI ethics. to better understand the problem and develop the questions of in-
The world largest technical professional organization, IEEE launches terest. The questions are finally developed based on the research
the guidelines Ethically Aligned Design (EAD) [7] that provides a concepts discussed in the mentioned research sources [3-9]. The
framework to address the ethical and technical values of AI sys- details of the research questions are provided in Section 1.
tems based on a set of principles and recommendations. The EAD
framework consists of the following eight general principles to 3.2 Data sources
guide the development and implementations of AI-based systems: The authors had a series of team discussions to identify the list of
human rights, well-being, data agency, effectiveness, transparency, digital data sources. The selected digital repositories are explored
Ethics of AI: A Systematic Literature Review of Principles and Challenges EASE, 13-15 June 2022, Gothenburg, Sweden

fifth authors, which are finalised by all the authors in the regular
consensus meeting (see Table 1).

3.5 Study selection

Identifying the
Problem
Selecting the
Primary Studies
Mapping Results to
Research Questions The search string discussed in Section 3.3 is used to explore the
selected digital repositories. The search process was initiated on
Specifying the
Assessing the Analyzing Threats
23rd December 2020 and ended on 5th February 2021. The search
Research Questions
Studies Quality to the Validity string retrieved total 811 studies in the first phase, which were
further filtered based on the study title, abstract and keywords
Defining the
Study Protocol
Analysing and
Synthsising the Data
Documenting the
Results
(see Figure 2). In the second phase of the selection process, the
inclusion/exclusion of the 60 studies are performed based on the
Research Protocol Review Results
full-text review. Finally, 24 primary studies are shortlisted using the
SLR approach. Moreover, backward snowballing [21] is performed
to search the references of the selected 24 studies. The backward
snowballing is previously used by Tingting et al. [1] to explore text
analysis techniques in software architectural and we used with the
aim to explore the references list of the selected primary studies
Figure 1: SLR research process
to identify relevant studies that are missed during the SLR process.
Additionally, 5 studies are selected, which are further filtered using
the inclusion/exclusion criteria (see Figure 2). Eventually, only 3
to extract the relevant data in order to address the given research
studies fulfil the selection criteria and the final data sets consist
questions (see Section 1). Finally, the following digital libraries
of total 27 primary studies (24 SLR + 3 backward snowballing).
are selected based on the authors SLR experience, discussions and
Final set of the selected studies is provided in Appendix A, where
guidelines provided by Chen et al. [4]: Springer Link, Science Direct,
each study is labeled as [Sn] to differentiate from the general list of
IEEE Xplore, Wiley Online Library and ACM Digital Library. These
references.
are the world leading digital data sources which collect a large
number of original information and communication technology
studies [4]. Literature Search String
("artificial intelligence ethics" OR "AI ethics" OR
String Execution
"machine learning
ethics" OR "software ethics") AND (“resistance” OR
3.3 Search strategy “barriers” OR
ACM Digital “limitations” OR “challenges”)
The research questions are analysed by the second and third authors Library

to extract the terms or keywords used for the search process. All the IEEE

authors participated in the group discussion to finalise the search Xplore

Digital Libraries

terms and retrieve the relevant data from the selected repositories.
Science
Pilot search terms and strings are made that finally contributed to Direct

develop the following agreed search string:

Spinger
("artificial intelligence ethics" OR "AI ethics" OR "machine learning Link

ethics" OR "software ethics") AND (“resistance” OR “barriers” OR Wiley Online

“limitations” OR “challenges”) Library

The “principles” and “guidelines” terms were excluded from the

final search string because these terms return irrelevant data from Figure 2: Studies selection process
different other domains. The given search string was specifically
testified during the pilot attempts to explore the data related to the
AI “principles” and “guidelines” and we noticed that it precisely
returns the desire results related to the RQ1, i.e., “principles” and 3.6 Quality assessment (QA)
“guidelines”. The assessment criteria are developed to evaluate the quality of the
The search terms are concatenated using “AND” and “OR” oper- selected primary studies and remove the research bias. The quality
ators to develop the search strings. The selected digital repositories assessment phase interprets the significance and completeness of
have a customised search mechanism. The search strings are exe- each selected primary study [9]. The QA criteria checklist provided
cuted using the personalised search mechanism of electronic data by Kitchenham and Charters [9] are analysed and designed the QA
sources. questions provided in Table 2. Each selected primary study evalu-
ated against the quality assessment questions (QA1-QA6). Score (1)
3.4 Inclusion/Exclusion criteria assigned if the study comprehensively addresses the quality assess-
The inclusion/exclusion criteria are developed to filter the search ment questions (see Table 2). Similarly, 0.5 points are assigned to
string findings and remove irrelevant, not accessible, redundant those who have partially addressed the QA questions. Studies with
and low-quality studies. The criteria are developed by the first and no evidence of addressing the QA questions are assigned 0 points.
EASE, 13-15 June 2022, Gothenburg, Sweden Khan et al.

Table 1: Inclusion/Exclusion criteria

# Inclusion criteria
In1 Consider articles that specifically focus on AI ethics.
In2 Primary studies published in conferences, research workshops, book chapters, journals and magazines.
In3 Peer-reviewed and available in full text
In4 Written in the English language
No. Exclusion criteria
Ex1 If two studies are published from the same project, then exclude the one with minimum contribution.
Ex2 Exclude grey literature material.
Ex3 Remove duplicate studies.
Ex4 Discuss ethics in other domains.

Table 2: Quality assessment criteria

No. Assessment Questions Score

QA1 Does the adopted research method address the research problem? 1/0.5/0
QA2 Does the study have clear research objectives? 1/0.5/0
QA3 Does the study explicitly discuss the proposed research approach? 1/0.5/0
QA4 Is the study clearly reported the experimental setting? 1/0.5/0
QA5 Do the study results and findings are systematically discussed? 1/0.5/0
QA6 Does the study present the real-world implications of the research? 1/0.5/0

3.7 Data extraction AI_Ethics <- [Link](AuthorsGroup1=c(3,5,

3,4,5,5,3,3,5,2,2,6,4,4,2),
The relevant data to address the RQs are collected by thoroughly
Author6=c(3,6,4,5,4,4,2,4,4,3,2,5,5,4,2),
reading the selected primary studies and extract the AI ethics prin- Author7=c(2,5,4,4,3,4,2,5,3,4,2,5,5,4,1))
ciples (RQ1) and challenges (RQ2). The extracted data are recorded KendallW(AI_Ethics, TRUE)
on excel sheets. Most of the data are collected by the second and KendallW(AI_Ethics, TRUE, test=TRUE))
third authors. They assess the quality of the primary studies based KendallW(t([Link][, -1]), test = TRUE)
on the criteria discussed in Section 3.6. Moreover, the first, fourth
and fifth authors participated in the review meeting to finalize the
QA score of each study (Appendix A). Additionally, authors at posi- Table 3: Kendall’s coefficient of concordance (W) test val-
tions six and seven are invited to assess the inter-personal bias in ues (DS-Data Set, df-Degrees of freedom, 𝜒2-chi square, S-
the data extraction process. They were asked to randomly select 15 Subjects, R-Raters, PV-Probability Values)
primary studies from a set of total primary studies and perform the
data extraction process as conducted by the other authors. Finally, DS 𝜒2 df S R PV W
the Kendalls coefficient of concordance (W) test is performed to sta- AI_Ethics 33.287 14 15 3 0.002 0.792
tistically evaluate the significant differences in the data extraction
process between both the groups (authors 1-5 and 6-7). Kendalls
coefficient of concordance (W) is a widely adopted statistical ap- 4 REPORTING THE REVIEW
proach to evaluate a level of agreement between groups of people,
The data collected from the selected 27 primary studies are analyzed
which assess a set of n-objects [16]. The Kendalls coefficient of
and discussed in the following sections.
concordance (W) assessment score varies from 0 to 1, where (W=1)
shows strong agreement between the people and (W=0) refers to
complete disagreement [16]. We used R-3.6.3 to conduct the (W)
4.1 Temporal distribution
test for evaluating the level of agreement between both groups The year wise distribution of the primary studies is shown in Figure
of authors. The test results (W= 0.792) given in Table 3 show that 3. Of the 27 studies, total 4, 17, 4 and 2 are respectively published
both author’s groups are at a positive agreement level for the data in 2021 (till 5th February), 2020, 2019 and 2018. The first relevant
extraction process. Based on the test results, we identified that no study was found in 2018 and since then, there has been a gradual
personal bias exists between the authors that could impact the data increase in the number of research publication. The SLR string was
extraction process. Following is the R-code developed to run the finally executed on 5th February 2021, therefore the given results
(W) test. only cover the first two months of 2021. The increasing number of
publications reveal that AI ethics is significant, and state of the art
library(DescTools) research direction. There is still need of substantial research work
to explore ethics in AI.
Ethics of AI: A Systematic Literature Review of Principles and Challenges EASE, 13-15 June 2022, Gothenburg, Sweden

Types of Publications
Book Journal Magazine Conference
Chapter Article Publication Proceedings

18
17
1
16
1 Conference
15 Proceedings
14 2 (06, 22%)
13
12 Magzine
11 Publication Journal
10 (01, 4%) Types of Articles
9 Publications (19, 70%)
8
7
13
Books
6
Chapters
5 (01, 4%)
4
1
3 Publication
2
2 Growth
1
1
3 Figure 4: Word cloud of the identified AI ethics principles
1 2

2018 2019 2020 2021 Years of Publications

Years and Type of Publications

range of system stakeholders; however, the level of transparency
should be varied for them [S4].
Figure 3: Temporal and publication types based distribution
of primary studies 5.1.2 Privacy. AI/autonomous system must assure user and data
privacy throughout the system lifecycle. It could broadly be defined
as “the right to control information about oneself” [S22]. Regulatory
institutions are consistently involved in establishing legislation
4.2 Publication type for data privacy and protection [S7]. However, privacy becomes
The selected primary studies are classified across four major types more challenging in data driven AI environment, where the system
i.e., journal, conference (including workshop), book chapter and subsequently processes user data including cleaning, merging, and
magazine. Figure 3 shows that 19 (70%) studies are published in interpretation [S7]. The data access in self-governing AI systems
journals, 6 (22%) in conferences, 1 (4%) book chapter and 1 (4%) is develop the primary concern of data privacy, which is commonly
a magazine article. We noticed that journals are the most active related to security and transparency [S21]. It is worth noting that AI
venues to publish relevant studies. technologies bring complex challenges associated with data privacy
and integrity, which demand more relevant future research [S22].
5 DETAIL RESULTS AND ANALYSIS
5.1.3 Accountability. Accountability is the third most frequently
The detail results to address RQ1 and RQ2 are discussed in the
reported principle which specifically focuses on liability issues
following sections.
[S5]. It refers to safeguard justice by assigning responsibility and
prevent harm [S3]. The stakeholders must be accountable for the
5.1 RQ1 (AI Ethics Principles) system decisions and actions to minimize the culpability problems
The final set of the primary studies consist of 27 articles and total 21 [S4, S5]. Ensure both technical and social accountability before
AI ethics principles are extracted from these articles. The identified and after the system development, implementation and operation
principles along with their respective references are provided in Ta- [S5]. Accountability is closely linked with transparency because the
ble 4. Moreover, a word cloud is generated to graphically represent system must be understood before making the liability decisions
the significance of the reported principles (See Figure 4). Of the 21 [S5].
principles, transparency (n=17) is the most frequently mentioned
principle, followed by privacy (n=16). The third and fourth most 5.1.4 Fairness. Fairness is considered a significant principle of AI
common principles are accountability (n=15) and fairness (n=14) ethics. Discrimination between individuals or groups made by the
respectively. decision-making systems lead to ethical fairness problems, which
impact public values including dignity and justice [S11]. Avoiding
5.1.1 Transparency. Transparency of operations is a major concern unfair biases of AI systems could foster social fairness. AI and
in AI/autonomous systems [S5]. It answers how and why a specific autonomous systems should not deceive people by impairing their
decision is made by the system and further triggers the other con- autonomy [S4]. It could achieve by explicitly making the decision-
structs including interpretability and explainability . It should not making process more transparent and identifying the accountable
only consider for the AI system operations, but must be part of the entities.
technical process [S5] to make the decision-making actions more Analysis. Based on the SLR findings, we identified that the above
transparent and trustworthy. Both operational and technical trans- principles received significant attention, which are compatible with
parency could be achieved by developing standards and models the widely adopted accountability, responsibility and transparency
that measure and testify the levels of transparency. Such standards (ART) framework [6] of ethics in AI. Responsibility is not a highly
could assist the AI system development organizations to assess cited principle in the selected primary studies and the reason might
their level of transparency and provide best practices for further be that it is considered an associated one with accountability [15].
improvements. Moreover, transparency should consider for a wide Moreover, Vakkuri et al. [S5] developed a relational framework
Ethics of AI: A Systematic Literature Review of Principles and Challenges EASE, 13-15 June 2022, Gothenburg, Sweden

organizations are reluctant to adopt these principles which are Table 5: AI ethics challenging factors
highly vague in their definition [S23]. For example, it is not clear
how specifically consider “fairness” and “human dignity” in AI # Challenges Reference
ethics [S17]. It is very challenging to consider AI ethics in real Ch-01 Lack of ethical knowledge [S14], [S15], [S18], [S21], [S23]
world settings using these vaguely formulated principles [S3]. Ch-02 Vague principles [S3], [S17], [S24]
Ch-03 Highly general principles [S19], [S20]
5.2.3 Highly general principles. The available principles are highly
Ch-04 Conflict in practice [S19], [S16]
general and broad in concept to specifically consider in the AI Ch-05 Interpret principles differently [S19], [S20]
industry [S18]. They are subjective in the term and used in various Ch-06 Lack of technical understanding [S10], [S13]
other domains than AI. Policymakers involved in drafting AI ethics Ch-07 Extra constraints [S11], [S24]
principles might not have strong technical understanding of AI Ch-08 Lack of Audit and Monitoring [S20]
system development processes, which makes the principles more Ch-09 No Legal frameworks [S20]
general and ambiguous. Ch-10 Business interest [S25]
Ch-11 Pluralism of ethical methods [S14]
5.2.4 Conflict in practice. Organizations, committees and groups Ch-12 Cases of ethical dilemmas [S14]
involved in developing the AI ethics guidelines and principles have Ch-13 Machine distortion [S14]
opinion conflicts regarding the real world implementation of AI Ch-14 Lack of Guidance and Adoption [S26]
ethics [S13, S16]. For example, the UK house of lords suggested that Ch-15 Lack of Cross-cultural Coopera- [S27]
robots cannot solely be in operation, but they should be guided by tion
human beings [S10], on the other hand, in various hospitals’ robots
make autonomous decisions in diagnosis and surgical endeavours.
It shows interpretation and understanding conflict for AI ethics in
practice. We noticed that very few studies are published where the barri-
ers of AI ethics are directly or indirectly mentioned. It is evident
5.2.5 Interpret principles differently. AI ethics principles are widely
from the frequency distribution of the challenging factors given in
considered ambiguous and general by majority of the organizations
Table 5. This finding reveals that the AI ethics challenges aspect
[S20]. It has been found that tech firms involved in the development
is very young field and requires considerable research effort from
of AI and autonomous systems follow ethical guidelines based on
diverse disciplines to be mature. The significance of AI technologies
their own understandings [S27]. There are no universally agreed
in various sectors calls for rush research to uncover the relevant
ethical principles that can bring all the institutions on one page.
challenges that hinder the process of considering ethics in AI.
5.2.6 Lack of technical understanding. The policymakers have lack
of technical knowledge, which makes AI ethics in practice a chal- Key Findings of RQ2
lenging effort [S10, S13]. They are not aware of the technical aspects
of AI systems and the advancement in AI technologies as well their Finding 3: We noticed that only (n=17, 63%) primary stud-
limitations. Lack of technical understanding develops the gap be- ies discussed the AI ethics challenging factors.
tween system design and ethical thinking [S10]. The ethicists must Finding 4: The core themes of the identified challenges
have skills of grasping technical knowledge using their ethical are: (knowledge and expertise, organizational management,
framework [S10]. tools and technologies).
Finding 5: Lack of ethical knowledge (n=5, 29%) is the most
5.2./ Extra constraints. Situational constraints could interfere with frequently cited challenge followed by Vague principles
an employee’s motivation to consider ethics in AI system devel- (n=3, 18%).
opment process. The possible constraints include lack of informa-
tion, incorrect guidelines, organizational rules and policies, and top
management interruptions. Organizational management should be The long-term plan of the research is to propose a maturity
highly aware of the AI ethics debate and keen to engage with the model that could be used to evaluate the ethical capabilities of the
constraints proactively [S26]. organizations involved in developing AI systems. The findings of
Analysis. The above reported challenges provide an overview this systematic review are the initial inputs for the development
of the most common and frequently cited factors that could be of the proposed model. Figure 7 shows the preliminary structure
potential barriers for scaling ethics in AI. Lack of ethical knowledge of the model and demonstrates how the findings of this review
is identified as the most common challenge of AI ethics. Major contribute in the development of the principles and challenges com-
ethical mistakes are made because of no moral awareness of specific ponent. The identified principles and challenges will be classified
problem [S14]. Practitioners only consider software development across capability and maturity levels. Moreover, best practices will
activities as the main responsibilities; however, they have limited provide to tackle the identified challenges and implement the AI
interest to consider ethical aspects [S5]. The ethical uncertainty in ethics principles. The given model is a proposed idea that will be
AI systems could only be diminish by acquiring ethical knowledge. systematically developed based on the industrial empirical stud-
Continuous awareness of ethical policies, codes and regulations ies and the concepts of the widely adopted CMMI process model
assist to properly manage the ethical values in AI and autonomous [17]. Case study approach is selected to evaluate the real-world
systems. significance of the model.
 Journal of Digital Economy 1 (2022) 44–52

Contents lists available at ScienceDirect

Journal of Digital Economy

journal homepage: [Link]/en/journals/journal-of-digital-economy

Ethical governance of artificial intelligence: An integrated

analytical framework
Lan Xue, Zhenjing Pang *
School of Public Policy and Management, Tsinghua University, Beijing, 100084, China

A R T I C L E I N F O A B S T R A C T

Keywords: Emerging technologies have faced ethical challenges, and ethical governance has changed over
Artificial intelligence time managing these technologies. The governance paradigm has gradually changed from scien -
Ethical governance tific rationality to social rationality and ultimately to a higher ethical morality. The trend of
Autonomous driving
seeking higher levels of ethics and morality provides a rich theoretical underpinning for the ethical
governance of artificial intelligence (AI), which is a complex and comprehensive project that in-
volves problem identification, path selection, and role configuration. Ethical problems in AI can
also be identified in technology, value, innovation, and order systems. In the four major systems,
the basic patterns of ethical problems can become uncontrolled risks, behavioral disorders, and
ethical disorders. When considering the path selection, AI governance strategies such as ethical
embedding, assessment, adaptation, and construction should be implemented within the tech -
nology life cycle at the stages of research and development, design and manufacturing, experi-
mental promotion, and deployment and application, respectively. Looking at role configuration,
multiple actors should assume different roles, including providing ethical factual information,
expertise, and analysis, as well as expressing ethical emotions or providing ethical regulation tools
under different governance strategies. This study provides a comprehensive discussion regarding
the practical applicability of AI ethical governance using the case of autonomous vehicles.

1. Introduction

Since the 1st Industrial Revolution, disruptive innovations have created a series of purposeful, and irreversible progresses that have
led us to the 4th industrial revolution through which a remarkable set of breakthrough innovations have been introduced such as genetic
engineering, new materials, new energy, Internet technology, and artificial intelligence (AI). These breakthroughs have been the advent
of an Axis Era of human technological revolution. While people are accepting the emerging technologies’ empowerment, at the same
time, they are actively constructing a system of governance for these technologies to avoid falling into the trap of what Heidegger refers
to as technological Gestell or fetishism. People have remained vigilant regarding the technological leap that crosses over the blurred
boundary between technology, human, nature, and society.
AI is the most typical of these technologies. People have been exploring AI from conception to application since the term of AI was
created in 1950s. With the increased computing power, availability of big data, and advances in algorithm, AI has permeated to the
entire production process of knowledge, technologies, and products. In particular, AI has become the core driving force for the digital
and economic transformation with the application of ANN-based deep learning using data gathered through AI, algorithms, full

* Corresponding author.
E-mail address: pangzhenjing@[Link] (Z. Pang).

[Link]
Received 27 June 2022; Received in revised form 7 August 2022; Accepted 14 August 2022
2773-0670/© 2022 The Authors. Published by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC
BY-NC-ND license ([Link]
L. Xue, Z. Pang Journal of Digital Economy 1 (2022) 44–52

coverage computing power, efficient replication, and multi-source heterogeneity. This process opens the gate to the smart era.
At the level of knowledge, technology, and application, the expansion of AI is characterized by strong penetration, high complexity,
and technological breakthroughs. While AI promotes the convergence of multiple elements, enhances the interaction of multiple
subjects, and facilitates the fusion of multiple states, the progress in AI research and application has also influenced the existing ethical
and moral order. People must walk the fine line between welfares generated by AI and the ethical risks brought by AI. Unlike other
emerging technologies, AI is characterized by a set of tangled attributes, such as hidden technical core, anthropomorphic technical form,
opportunities for cross-domain application, intertwined interest subjects, multi-dimensional technical risk, and complex social impacts.
These attributes generate many ethical concerns in the development and application of AI technology, including problems related to
infringement, discrimination, problems associated with technological leviathan, digital divide, information cocoons, the Matthew ef-
fect, and other issues.
Historically, discourse on these issues was dominated by advanced countries where technologies were first developed and the
challenges were first encountered. Various ethical initiatives, declarations, and rules were usually launched from these countries.
However, with the progresses made by China and other emerging countries, the ethical challenges of new technologies are recognized
and encountered by a much wider part of the global community, including China, which can no longer stay behind. China has to join
these discussions and contribute to the debates in a constructive way. In this spirit, our study tries to provide an integrated analytical
framework for the ethical governance of AI and to identify a practical path for its application. The rest of this paper is structured as
follows, section 2 reviews the past studies that explore ethical governance in emerging technologies and AI; section 3 provides inte-
grated analytical frameworks for the AI ethical governance; section 4 discusses the case of autonomous vehicles, section 5 concludes the
paper.

2. Literature reviews

2.1. Ethical governance in emerging technologies

Following the theoretical origins, ethical concerns in emerging technologies have followed a clear line of progression. In the 1960s,
the emerging new technologies such as nuclear energy and chemical-intensive industries caused prominent environmental problems.
With these problems, governments formulated policies regarding regulations to safeguard the ethical use of these technologies.
Technology assessment became the core approach to technology governance (Baram, 1973). Thereafter, an expert decision-making
model (technocracy) was confirmed by the legitimacy of knowledge, in which political experts with policy experience and technical
experts with knowledge authority worked collaboratively to make institutional arrangements. These experts chose technology gover-
nance tools by assessing the impact based on predictable paths of technology development (Sarewitz, 2011). In the 1980s, studies on
genetics leaped from theoretical knowledge to technological application, thereby, opening a potential Pandora's box for artificial life.
This led to a discussion of the uncertainty and ethical dilemmas of emerging technologies.
A series of short-term technological disasters, such as the European mad cow disease crisis and the Chernobyl Nuclear Power Plant
accident, further dissipated trust in public institutions regarding technological decision-making, which led to the gradual decline of the
regulation and assessment-based governance model. The precautionary approach emerged as a new method to govern the emerging
technologies, in which post-normal science was described as deficits in knowledge and information as well as burgeoning ethical tension
(Stirling, 2016). This new approach advocates an active precautionary policy framework until the ethical dilemmas in emerging
technology open discussion from multiple perspectives. With the commercialization of genetically modified crops, the theory and
practice of the governance of emerging technology centered on the precautionary principle. In the 1990s, the ethical, social, and
economic implications of technology became the focus of discussion with the implementation of the Human Genome Project. Then, the
ethical, legal, and social implication (ELSI) model emerged as an emerging corrective mechanism for technology governance to fill the
humanistic gap. The ELSI model advocated the inclusion of broader ethical values, legal, and socio-economic implications of technology
(Michael, 2008). Although the ELSI model has moved beyond the initiation stage, it was not promoted in the practice of emerging
technology governance.
At the beginning of the 21st century, the nanotechnology breakthrough became a precursor of the fourth technological revolution,
and the concept of anticipatory governance was proposed based on the reflections of ELSI. The core of anticipatory governance
embedded social values, ethics, and public preferences into the scientific research process (Guston, 2010) to shape technologies from an
early stage and make them more ethical (Rip et al., 1995; Guston, 2002; Grin, 2000; Wynne, 2002; Friedman et al., 2002). Moreover,
anticipatory governance formulated open scientific research and contributed to the development of an ontological aspect of the
co-evolution of technology and society. Since nanotechnology has failed to facilitate industrial renewal in a revolutionary way, people
have turned to the governance and innovation of emerging technologies. As a result, the concept of responsible research and innovation
(RRI) emerged in 2010 and was adopted by the European Union (EU) 2020 Framework Program. With the application of synthetic
biology research, RRI has become the mainstream paradigm for emerging technology governance in the EU (Owen et al., 2012). The RRI
emphasizes the paradigm shift of technology governance from traditional risk-based — such as assessment of technological, ethical,
legal, or social impact, the precautionary principle, and anticipatory governance—to a paradigm of shaping the responsibilities of
scientific researchers. The paradigms also shift to science and technology innovation, the institutional responses to innovation, the
reshaping of public responsibility for science development, and the establishment of public participation in science.

45
L. Xue, Z. Pang Journal of Digital Economy 1 (2022) 44–52

2.2. Ethical governance of AI

The ethical concerns in emerging technology governance provide rich theoretical resources for the ethical governance of AI. The
current studies on AI ethical governance focus on three specific research areas: concept, framework, and subject. At the conceptual level,
different perspectives have been used to determine the types of AI to be developed and influence the application process through the
definition of core concepts, objectives, and values. Although the concept of AI is still controversial, it does not affect the application of AI
from a governance perspective. The core conceptions that influence human-technology relations include “Beneficial AI” (Friedman
et al., 2002), “Ethical AI” (UK House of Lords, 2017), “Trustworthy AI” (OECD, 2019), and “Responsible AI,” which have been proposed
by China's New Generation AI Governance Committee. These concepts guide stakeholders to think about the direction of AI develop-
ment, while the interpretation of the specific connotations and the resulting requirements for products and behavioral norms also guide
the governance of ethical issues related to AI.
At the framework level, a holistic analytical framework was proposed to localize and modularize the complex issues related to the
main ethical concerns in AI. Such research focuses on the origins, innovations, improvements, and expansions of emerging technology
governance theories. In addition, this area of research embeds AI ethical issues in the existing theoretical frameworks or extracts useful
elements from existing theories and maps them into the analytical domain of AI ethical governance to provide paths for ethical issues
arising from the emergence, application, and development of AI. The goal aims to provide a pathway for solving ethical problems and
exploring existing governance frameworks. Typical examples include accountability and explainable AI, which focuses on issues such as
responsibility; discrimination of data mining, which focuses on issues of fairness; privacy by design, which focuses on privacy.
At the subject level, research focuses on practical rationality to delineate the scope of the subject matter, toolsets, and path selection
involved in AI ethical governance. The major focus is on the relationship and structure of responsibility and power distribution among
different stakeholders in the process of AI ethical governance. A holistic AI ethical governance framework should include technical,
organizational, and policy levels.
Recently, many new ideas have emerged in the discussion of governance mechanisms and emerging technologies frameworks.
Moreover, a more complete theoretical framework has gradually been formulated that provides an in-depth analysis of the power re-
lationships among different actors, such as government regulators, enterprises, and the public regarding the regulation of emerging
technologies. Tentative and adaptive governance models are empirically adopted in exploring emerging technologies, considering the
dynamics, uncertainties, innovations, and potential risks. In addition, the constraint and incentives between regulators and the regu-
lated ones are formed based on the grassroots discretionary power of front-line regulators to achieve more effective self-restraint.
Meanwhile, these discussions have become the theoretical basis for innovation in AI ethical governance mechanisms.
In summary, ethical concerns in the governance of emerging technology have followed a clear trajectory of paradigm change, from
the assessment of technology to the precautionary principle; from ethical, social, and legal assessment to anticipatory governance; from
responsible research and innovation to adaptive or tentative governance. Each paradigm change reflects a trend from scientific ratio-
nality to social rationality, with a strong focus on ethics and morality. This provides a rich theoretical underpinning for the ethical
governance of AI in the smart era. Existing research has formulated a systematic research outline, which is of great significance in
guiding the practice of AI ethical governance. However, this line of research failed to probe profoundly the problematic nature of ethical
governance of AI and analyzed the different ethical issues, using empirical data, model training, verification, application evaluation, and
feedback. In addition, the existing studies have not examined these issues based on the life cycle of AI development from R&D to
application. Moreover, they have not put forward any corresponding adaptive governance solutions or portrayed the role of different
actors (e.g., technical actors, policy actors, social actors) in the ethical governance of AI. This study tries to fill these gaps by constructing
an integrated framework for AI governance, recognizing the problem, application scenarios, and role configuration, and clarifying the
practical path for the ethical governance of AI for the autonomous driving scenario as a useful exploratory case.

3. Integrated analytical frameworks for the AI ethical governance

Recognizing the tremendous potential and the clear risks, more than 70 programs on the ethical principles of AI have been proposed
by different national and regional governments, intergovernmental organizations, scientific research institutions, non-profit organi-
zations, scientific and technological societies, and enterprises globally. These proposals focused on 10 major themes: human-
centeredness, cooperation, sharing, fairness, transparency, privacy, external security, internal security, accountability, and long-term
applications. On November 24, 2021, UNESCO released a Recommendation on the Ethics of Artificial Intelligence, which was the
first global normative framework for the ethical governance of AI. The recommendation identifies 10 principles and 11 areas of action
for regulating AI technologies. The recommendation also proposes that the development and application of AI should reflect four major
values: (1) respect, protect, improve human rights and enhance human dignity, (2) promote the development of the environment and
ecosystems, (3) promote diversity, inclusion, and equity in the workplace, and (4) build a peaceful, just, and interdependent human
society.
However, translating these ethical principles in practice is a complex and challenging endeavor and need a more systematic approach
in identifying problems, selecting solution paths, and assigning roles to relevant stake holders. In the following discussion, we will first
identify potential problems associated with AI application from the perspective of the social system where such application occurs.
Second, we will try to delineate paths to ethical use of the technology based on the technology life cycle. Finally, we will construct a role
configuration that can lay out relevant responsibilities to multiple stakeholders involved in the development and application of AI.

46
L. Xue, Z. Pang Journal of Digital Economy 1 (2022) 44–52

3.1. Problem identification

Fremont E. Kast used systems theory to analyze management issues comprehensively. He indicate that all management problems
arise from goal-orientation system, social-psycho system, an integration system of structured activities and technical systems. According
to social systems theory, the ethical governance of artificial intelligence is rooted in four major systems (Fig. 1). The first is the tech-
nology system, comprising the knowledge and artifacts of technology, which manifests as products, hardware, software, and services of
AI, and the ethical governance emphasizes effective control of technology risks. The second is the value system, focusing on technology
assessment, expressed as people's moral rationality regarding the value orientation. From the perspective of ethical governance, AI is
incapable of breaking the normal relationships between technology, people, society, and nature, but it can have an influence on AI's
ethical acceptability. The third is the innovation system, in which the socialization of AI technology is manifested as innovative be-
haviors and activities under different application scenarios. Here, the ethical governance dimension focuses on the constraint of
innovative behavior for AI. It also emphasizes that the instrumental rationality of behavior cannot exclude considerations on ethical
responsibility. The fourth is the order system—the system for distributing rights, powers, interests, and responsibilities under technical
conditions, expressed as the social impact brought about by the application of AI. Here ethical governance emphasizes that the dis-
tribution of rights and responsibilities, power structures, and interest patterns among relevant stakeholders are embedded in social
scenarios to maintain a stable social order.
In the technological system, the ethical problems of AI are endogenous and manifest as the probability of uncontrolled risk, man-
ifested as the inability to predict, explain, calculate, evaluate, and control technical risks scientifically. It is also presented as the inability
to dissipate the negative effects of AI applications. The uncertainty, bias, and vulnerability of AI technology are the root cause of all these
ethical problems. Ethical problems arise when the black-box algorithms, their self-reinforcing character, and the proliferation of al-
gorithms are fed with biased data. When machine learning heavily relies on code and data samples, stability can hardly be guaranteed.
Moreover, multiple repeated runs of sample data can mislead the machine to make wrong ethical decisions due to misleading as-
sumptions. For example, discrimination is often a by-product of the unpredictable and unconscious results of algorithms. Algorithmic
engineers make conscious choices and face challenges in identifying AI-related ethical problems.
From the perspective of the value system, the ethical problem of AI has a relational character, which is a disorder in the ethical
relationship, revealing that the development of AI boundaries is blurred when considering the relationship between humans and
technology, as well as the relationship between technology and society. Ethical themes before launching AI are different from those
ethical themes in the modern era of AI. Formerly, the scope of discussion was limited to the relationship between humans. Presently, the
scope of discussion now extends to the relationship between humans and machines, leading to the formulation of machine ethics, where
the autonomous system is capable of posing risks to a human. The question is how do human beings view their relationship with
intelligent machines? Generally, only humans have morals, and their biological senses determine their ethical behavior and make them
moral subjects since only human beings can rationally think and communicate. With the development of AI, moral status related to AI
products is the focus of discussion. The relationship between humans and intelligent machines is different from the relationship between
humans, probably changing the position of human morality. Thus, defining the ethical relationship between humans and machines is
challenging.
In the innovation system, the potential ethical problem AI's development is that the innovative practices and applications of AI
excessively pursue the means to reach an end, while ignoring the values and beliefs. Thus, AI is a used as a vital strategic resource for
promoting the development of science and technology, industrial optimization, upgrading, and productivity without proper consid-
eration of the ethical implications in using AI. The comprehensive progress of AI in data, algorithms and computing power has become
the core driving force for the digital transformation of the economy and society. If we pay too much attention to the tool value and ignore

Fig. 1. Problems identification in the ethical governance of artificial intelligence.

47
L. Xue, Z. Pang Journal of Digital Economy 1 (2022) 44–52

the moral responsibility of AI in the process of technological innovation, the development of AI will be unsustainable.
In the order system, the ethical problem of AI is derivative, leading to the disorder of ethical outcomes. In the process of AI so-
cialization, the reconfiguration of rights, power, interests, and responsibilities are caused by differences in cognition, understanding,
and usability between different groups or individuals, without institutional mechanism to function as a corrective measure to maintain
fairness and justice. The development of human society lies in the pursuit of harmony and the maintenance of social justice, which has
been the basic orientation of social ethical values. However, the development of AI has dismantled the social division of labor and
overturned traditional labor relations, causing significant structural unemployment and affecting social justice. With the breakthrough
of AI applications, privacy infringement, algorithmic discrimination, the digital divide, and blurred responsibility have generated many
legal and ethical challenges.

3.2. Path selection

Although AI technology itself does not have an ethical and moral quality, developers can incorporate ethical values into AI through
algorithm design, data selection, and model optimization. Thus, comprehensively understanding of ethical governance of AI is vital to
embed ethical considerations in each stage of innovation and application, involving adopting different regulatory mechanisms based on
different stages of technological development. Corresponding regulatory tools must be integrated at each stage of AI research and
development, design and manufacturing, and experimental promotion and deployment to prevent the probability of uncontrolled risk
and disorder in ethical relationships, behavior, and outcomes brought about by AI technology (Fig. 2).
The first stage is to embed an ethical model in the research and development phase. In this stage, ethical embedding involves
integrating AI technologies with ethical norms, comprising four major aspects. First, an analysis of the ethical goals embedded in AI
should be conducted through the basic theoretical assessment of the ethics of science and technology. This stage defines the stakeholders
and clarifies the value of rationality, normative conflicts, and human subject status. Moreover, the risks associated with R&D should be
clarified. In the second stage, the ethical goals of AI are implemented into specific architectures, which include physical carriers, al-
gorithms, and interfaces. Third, AI technological development is integrated with application scenarios to predetermine the specific value
norms. Fourth, the rationality of AI in terms of potential ethical norms is evaluated to determine whether the technology can reasonably
meet expectations.
The second stage incorporates ethical assessment in the design and manufacturing phase, involving the use of AI theory or tech-
nology to implement relevant activities to form systems, products, or services to meet specific needs. At this stage, the assessment was
oriented toward the future of the technology and ethical perspective. Rather than determining whether an AI technology is good or bad
and whether to accept or reject it, ethical assessment in the design and manufacturing phase should explore the new approach to
technology and consider the social functioning in the design process. In this phase, the ethical assessment is an analysis of potential AI
ethical problems. An ethical assessment entails anticipating and identifying potential risks, analyzing and clarifying ethical issues, as
well as developing solutions to ethical issues through peer review and communication with the government and the public.
The third stage involves ethical adaptations in the experimental promotion phase, entailing the use of AI systems, products, or
services within a certain range to observe the process of their integration with social systems, the AI socialization stage. At this point, AI
technologies have acquired specific social roles and passed the ethical assessment that should be integrated with social systems.
Considering the influence of social factors and adjusting societal value choices are vital to achieving social embedding and integrating
the AI product with the social value system. In the experimental promotion stage, ethical adaptations must be proactive. At the insti-
tutional level, the promotion of AI must be guided and amended through laws and regulations to follow the principles of fairness and
justice. AI products with negative potentials that run contrary to mainstream social values should be recalled. The establishment of a
social ethics education system is vital to the ethical issues of AI, advocating responsible innovation, and providing moral and intellectual

Fig. 2. Path selection: Ethical governance of artificial intelligence.

48