Principles Guiding the Use of AI in Arbitral Proceedings (Part II)

As outlined in Part I, a framework governing the use of AI in arbitration is rapidly taking shape. In this Part II, we explore how emerging principles—responsible use, human oversight, data protection, transparency, and fairness—interact with real-world arbitral practice. Where we set out relevant jurisprudence and rules, this is not intended to be an exhaustive survey, but instead to highlight recent practices that have emerged. As the use of AI and regulations continue to develop, there may be greater room for divergence across jurisdictions.

Balancing AI Efficiency with Human Oversight        

As the uses of AI in international arbitration expand, maintaining human oversight is key to ensuring that the use of AI remains efficient while complying with broader professional and ethical obligations.

For parties: Whether parties’ use of AI is protected by attorney-client privilege remains unsettled, and the law on this continues to develop. While we are not aware of any arbitral tribunal decision on the interaction between the use of AI and legal privilege, courts in the United States have started to grapple with this question, but clear principles are yet to emerge. In United States v Heppner (S.D.N.Y. 2026) the court held that communications with a publicly accessible AI tool were neither privileged nor protected as work product because they were not confidential and were not communications with, or documents prepared by, an attorney (see Blog post here). Conversely, in Warner v Gilbarco (E.D. Mich. 2026) the court held that a pro se plaintiff’s AI chats reflected her internal analysis and mental impressions generated using a tool, and were thus protected by the work product doctrine.

For counsel: A core risk associated with generative AI use is “hallucination”—where the tool produces plausible but incorrect citations, quotations, or summaries. In several instances, courts have criticized (and in some cases sanctioned) lawyers for filing submissions containing non-existent or inaccurate authorities (see e.g., the English High Court in Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank and the Singapore High Court in Tajudin bin Gulam Rasul and Mohamed Ghouse s/o Rajudin v Siriaya bte Haja Mohidden).

To mitigate this risk, guidelines increasingly recommend that counsel implement robust internal verification processes (including appropriate supervision), rather than relying blindly on AI outputs. The Singapore Ministry of Law’s Guide for Using Generative AI in the Legal Sector, for instance, recommends a risk-based approach: a “human-in-the loop” model for high-risk work products (where legal professionals review outputs and verify facts, citations, and legal reasoning), and a “human-on-the-loop” model for lower-risk tasks (where oversight is provided through supervised checks or structured sampling) (see Diagram 2). 

For arbitrators: AI-assisted summaries may influence how facts are selected, framed or prioritized. An increasing number of arbitral institutions, in their AI-specific guidelines to arbitrators, consistently emphasize the non-delegation of legal reasoning and decision-making to AI (see e.g., SVAMC Guidelines 2024, SCC Guide 2024, AAA-ICDR Guidance 2025, VIAC Note on AI 2025, and CIArb Guideline 2025).

The challenge is therefore not whether AI may be used, but how it can be integrated in a way that preserves efficiency gains while ensuring an adequate level of human oversight.

Divergent Approaches to Transparency          

The Illinois Supreme Court’s Policy does not require parties to disclose the use of AI in pleadings. That contrasts with the approach of one of the magistrate judges in the Northern District of California, which requires parties to disclose any part of their pleadings produced with the help of AI, as well as any AI-generated material used as evidence, failing which such material may be treated as inadmissible (see p 10). The underlying rationale is that the tribunal should be able to evaluate the tools and methods used to generate submissions and evidence.

Arbitral institutions have taken different approaches to disclosures by parties and arbitrators. 

• Disclosure by parties: Most of the guidelines do not require parties to disclose their use of AI. For example,  the SCC Guide and the AAA-ICDR Guidance (which applies only to arbitrators) are silent on disclosure by parties. The SVAMC Guidelines suggest that participants in the arbitration should consider disclosure on a case-by-case basis, taking into account the relevant circumstances, including due process and any applicable privilege (see Guideline 3). The CIArb Guideline requires parties to disclose their use of AI tools where this affects the evidence or outcome of the arbitration, or otherwise involves a delegation of an express duty toward the arbitrators or any other party (see Article 7.1).
• Disclosure by arbitrators: In contrast to the position on party-disclosure, most of the guidelines focus on circumstances when it would be appropriate for arbitrators to disclose the use of AI. The SCC Guide encourages tribunals to disclose “any use of AI” in research, interpretation of the facts and the law, or in their application. Guideline 7 of the SVAMC Guidelines considers disclosure to be appropriate when arbitrators rely on AI-generated information outside of the record. Under Guideline 9.1 of the CIArb Guideline, arbitrators are, by default, encouraged to discuss the use of any AI tool with the parties and provide parties with an opportunity to comment and oppose such use by the arbitrators. The AAA-ICDR Guidance mandates disclosure where the use of generative AI “materially impacts” the arbitration process or the arbitrators’ reasoning (see paragraph 4).

Ensuring Fairness in Data-Driven Systems     

While AI systems offer powerful analytical capabilities, their effectiveness depends on training data, design choices, and deployment context. AI tools trained on publicly available or jurisdiction‑specific datasets may underrepresent legal traditions, procedural norms, or linguistic nuances relevant to a particular dispute. That can affect tasks such as issue framing, expert selection, or comparative legal analysis. 

The UK Judiciary’s AI Guidance focuses on user awareness, whereas Brazil has adopted an affirmative duty to identify, assess, and address distortions that may arise from data or system design. 

An illustrative example is the initiative by the Chilean National Center for Artificial Intelligence, which is developing an open-source AI model trained on Latin American languages and cultural contexts: Latam-GPT. This project reflects a growing recognition that contextual and linguistic specificity can materially enhance the relevance of AI‑assisted outputs in legal and institutional settings.

At the same time, current experience confirms that no AI system can capture the full nuance of every factual or cultural context in which it is deployed. This raises an important practical question: to what extent can region‑specific models address limitations associated with training data and system design, particularly when applied to sensitive, judgment‑adjacent tasks? The issue is especially relevant in arbitration, for tasks like the selection of arbitrators, counsel, or experts. As highlighted in the SVAMC Guidelines, AI tools that are not appropriately designed or contextualized may influence outcomes in ways that warrant careful consideration from a fairness and representativeness perspective.

Confidentiality in an Era of AI    

The handling of confidential information in AI systems must be carefully aligned with the tool’s data-handling architecture, contractual safeguards, and security features. 

A critical implication is that superficial anonymization may not be enough. AI systems may, in certain circumstances, be able to re-identify parties by correlating case details with publicly available information. Large Language Models (LLMs) deployments must therefore be configured so that confidential client information is not used to train general models. Counsel should consider seeking “zero-retention” or “no-training” commitments clauses in service agreements with AI providers. Using 'enterprise or private LLMs—rather than public interfaces—is a practical necessity to ensure that sensitive data remains siloed.

The procedural framework of arbitration is also adapting. Under Rule 16.1(a) of JAMS’ AI Rules, a Protective Order applies automatically to all materials disclosed in the proceedings. The production and inspection of AI systems are limited to experts and conducted in a secure environment, with no removal of the materials—measures designed to safeguard confidentiality.

Conclusion

The focus in arbitration is no longer on whether AI tools may be used, but on how they can be integrated responsibly and effectively. When properly used, AI can enhance the conduct and efficiency of arbitration while remaining consistent with the fundamental principles of fairness, equality, and integrity that underpin the arbitral process.

*The authors thank Boyan Arshinkov for his research and initial draft of the post.