Q1: We will start with a sketch of the context: when we speak of AI/Tech in arbitration,

what exactly are we talking about? Louise to start followed by Chinaecherem and then
Salma.
The Organisation for Economic Co-operation and Development (OECD), defines an AI
system as a "machine-based system that, for explicit or implicit objectives, infers, from the
input it receives, how to generate outputs such as predictions, content, recommendations, or
decisions that can influence physical or virtual environments". This definition is similar to the
definition of AI in the EU Artificial Intelligence Act, 2024. It has also been adopted in the
CIArb Guidelines on the Use of AI in Arbitration, 2025. It emphasizes the autonomy and
adaptability of AI systems, which vary in their ability to operate independently and adjust
their behaviour after deployment.

In summary, AI and Tech in arbitration refers to the use Artificial Intelligence (AI) and other
technological tools to enhance efficiency and the accuracy of dispute resolution processes, by
performing tasks like document review, legal research, data analysis, case management, and
potentially even assisting with factual or legal analysis for decision-making. We are talking
about the deployment of technology, some of which are programmed to mimic human
thoughts and actions in learning and problem-solving, to automate tasks and streamline
workflows, leading to cost and time savings.

Q2: What core principles of arbitration do we think will be impacted by

AI/Tech? (Chinaecherem to start followed by Salma and then Louis)
1. Neutrality and fairness
2. Confidentiality and privacy
3. Diligence
There is also the issue of Legitimacy of decisions rendered by AI, which also raises the
question of liability.

Neutrality and Fairness

While AI outputs may give the illusion of neutrality, they are likely to have encoded hidden
structural biases. AI systems inherit biases from their underlying models, training processes,
and datasets. Even when datasets are balanced, the training process can still introduce or
amplify bias. Additionally, the model architecture and design choices, such as how it is
structured or what features it prioritizes, can embed further biases into the system. They may
embed prejudices linked to geography, legal traditions, or even language, eroding neutrality
in arbitral proceedings. For instance, an AI trained predominantly on Western jurisprudence
might undervalue informal contractual norms recognized in African, Middle Eastern, or Asian
contexts.
Although it has been argued that AI reduces human bias by applying consistent logic,
algorithmic bias in AI tools can subtly affect the objectivity and independence of arbitrators
by perpetuating pre-existing prejudices embedded in the data or algorithms, producing
skewed outputs. This situation risks undermining the fundamental requirement of impartiality
in arbitration.

Transparency and Accountability in Decision-Making

Traditional judicial decisions or arbitral awards involve the explanation of the reasoning
behind decisions. On its part, AI systems have obscure output system. They analyse and
process datasets to detect patterns and generate recommendations, without providing the
exact steps or logic behind their conclusions. Thus, how it weighs facts, interprets evidence,
or prioritizes information, are not transparent or easily understandable by the users. This
opacity in AI systems’ decision-making process has been termed the "Black Box" problem.
The black box problem means that arbitrators, parties, and counsel often cannot trace or
articulate the exact reasoning behind the AI’s output, which poses a challenge in legal settings
where reasoned decisions are required.
Human judgment, intuition, empathy, and discretion are critical components in adjudicating
many disputes. Decision-making often hinges on interpretation of witness demeanour,
cultural norms, the spirit and intendment of an agreement or law, etc., all of which, I think, as
of this moment, are dimensions beyond the AI’s ability. AI’s inability to process moral and
contextual nuances of human disputes is a significant limitation to its effectiveness.

Confidentiality (Data Privacy & Security)

Confidentiality remains a cornerstone of arbitration. Feeding pleadings, witness statements,
or awards into third-party AI systems raises risks of data exposure, storage misuse, and
cyberattacks. Sensitive documents require absolute protection. The use of AI platforms
introduces the risk of unauthorized access, breaches, or misuse. AI provide quick fixes but
comes with the risk of accidentally disclosing sensitive data, making the owners susceptible
to cyberattacks and/or intellectual property theft. Every interaction with chatbots results in
storage of these data for use in further training of these tools. This raises significant questions
about data ownership, control, and trust in the arbitral process.
Open-source AI models (e.g. LLaMA2 for text generation; GPT-Neox and GPT-J for text
generation and instruction following, Whisper for speech recognition, etc.) pose particular
dangers, as parties cannot always know how inputs are stored, repurposed, or disclosed to
others. Robust protocols are therefore essential, including confidentiality agreements, vetting
of AI tools, and continuous cybersecurity vigilance as arbitration shifts away from hard-copy
submissions. It has also been suggested that another solution may be for arbitration bodies to
download, fine-tune, and deploy open-source AI models to local servers. This can run entirely
offline on a secure, local infrastructure.

Diligence
Arbitrators owe parties the duty of diligence, however, arbitrators relying on AI risk
affirmational authority bias, that is, the tendency to defer to machine-generated outputs. It has
been noted that constant or excessive delegation of reasoning to machines threatens the
human element of arbitration – judgement, discretion, and cultural sensitivity. There is also
the risk of developing cognitive inertia that may dull independent judgment.
While AI can lighten administrative burdens, the imperative and reasoning functions of
arbitrators cannot be delegated – AI must be restricted to supportive functions, never final
reasoning. As one commentator observed: “An arbitrator must know their case better than
anyone else. AI cannot replace that fundamental duty.”
To maintain credibility of the process, responsibility for arbitral outcomes must remain with
the arbitrator who must verify AI outputs and maintain control over all substantive decision-
making. AI must remain an assistant, not the decision-maker. Preserving the integrity,
legitimacy and enforceability of arbitration depends on maintaining this boundary.

Legitimacy
Apart from undermining the fundamental arbitration requirement of impartiality, the use of
AI for arbitral decisions raises the critical question of the legitimacy of such decisions. And
so, the question has been asked, when an arbitrator or arbitrary panel relies on AI for their
award, can that award be said to have been given by the arbitrator? Is it fair on the parties
who select arbitrators relying on certain criteria and qualities, to have their decision delegated
to a machine?
It may be argued that AI is an assistant, and once the arbitrator assumes paternity over its
output, it must be taken as the arbitrator’s. However, the fact must be acknowledged that
algorithmic bias in AI tools used in arbitration can subtly affect the objectivity and
independence of arbitrators by perpetuating pre-existing prejudices embedded in the data or
algorithms.
Also, while AI can improve efficiency by organizing data and assisting in drafting, as well as
the other uses highlighted above, there is the real risk that over-reliance on AI would
compromise the arbitrator’s independence and human judgement, which are essential for due
process. Besides errors, hallucinations, or inaccurate information generated by AI tools, that
may lead to flawed decisions (unenforceable awards), the question of the legitimacy and
credibility of decisions given by AI persists.

Liability
Another issue that remains a critical unresolved challenge is liability concerns. It is unclear
who bears responsibility or potential legal liability if AI tools produce erroneous, misleading,
or biased results that materially affect arbitration outcomes. Whether liability should rest with
the developers of AI software, the arbitrators who rely on the AI, or the parties who consented
to the AI’s use is not yet defined by law or institutional rules. This ambiguity creates
exposure risks for arbitrators and parties, discourages reliance on AI, and calls for clear
accountability frameworks to manage damages arising from AI failures or misuse. Until these
liability and regulatory issues are addressed, AI use in arbitration will require cautious
implementation and robust human oversight.

Q3: What are the possible risks we see and will see? Salma to start followed by Louise
and Chinaecherem.
1. Data Breaches and Unauthorized Access to sensitive data
2. Unintended data storage and replication
3. Hacking and System Intrusions (phishing attacks, the PCA of 2015 – in July 2015, the
website of the Permanent Court of Arbitration (PCA) was hacked in the midst of an
ongoing maritime border dispute between China and the Philippines. Malware was
implanted on the PCA's website which infected the computers of visitors, potentially
exposing them to data theft.
4. Evidence Tampering and deepfakes (this includes concerns as to witness coaching and
authentication challenges)
5. Ransomware and Denial-of-Service (DoS) Attacks
6. Insider Threats and Human Error
7. Hallucination: AI has been known to hallucinate; I think that is one attribute that has given
it notoriety. (Cite the case of administrative law authority and the search about the Past
NBA GS. Also, yesterday my boss sent me a news article, about a Singapore lawyer who
was, just last week, ordered by the court to pay S$800 (USD$620) in personal costs to the
other party, for citing a fictitious legal authority produced by generative AI.
Therefore, arbitrators must retain full responsibility for decisions and should critically
assess AI outputs rather than accept them uncritically. This involves scrutinizing the
underlying data sets used to train the AI, evaluating whether any debiasing measures were
applied, and ensuring that human judgment overrides any flawed or biased AI conclusions.
8. Sycophancy: One problem that have been identified with OpenAI Models is their tendency
towards sycophancy. In April this year, a user posted on the social media, “AI becoming
too sycophantic? Noticed Gemini 2.5 praising me instead of solving the issue” Many users
have noted the tendency of AI models to adjust their responses to align with users' views,
which has been nicknamed the digital yes-man. Earlier in April, ChaptGPT announced
that it was rolling back a particular update, the GPT-4o, because it was overly flattering or
agreeable—often described as sycophantic. Notice the problem was not that it was
flattering and agreeable, but that it was OVERLY so. That tells you what you need to
know. Beyond being irritating, this flattering nature of AI chatbots poses more serious
problems, as it raises serious concerns as to accuracy.

Q4: Do we see any positives/opportunities? Louise to start followed by Chinaecherem and

Salma
 Speed and efficiency
 Effective case management system
 Virtual hearings which reduce the risks and logistical nightmares associated with
travelling and invariably reduces carbon footprints.
 Secure and tamper-proof evidence storage and handling.
Adopting AI to improve operations will, in addition to improving efficiency, give
competitive advantage over other practitioners that fail to fully integrate AI into their
processes.

Q5: How will all of these impact on Africa: positive/negative: Salma to start followed by
Chinaecherem and Louise.
Cite the instance using the video shared by Chizaag
Machine learning requires expending funds; dedicated servers, bespoke virtual hearing suites,
closed source ai, etc. all require significant costs that may be regarded as prohibitive.

Q6: Any advise for African practitioners? Louise to start followed by Chinaecherem
and Salma. (30 seconds each)
Technological innovations undoubtedly open the door to new risks and introduce
vulnerabilities that tribunals and arbitration practitioners must now deal with. The
deployment of AI invariably involves significant challenges and costs, and there is no running
away from AI and digitisation and all forms of technological innovation. It is the new normal.
The days of ‘should we accept AI’ are long gone.
- We must welcome these innovations
- But let our approach be that of informed scepticism