Introduction

Generative AI has changed what arbitrators can do in seconds: summarize a voluminous record, prepare a first draft of the procedural history section of an award, or stress-test legal reasoning. The speed of that change has outpaced ethical frameworks, but the gap may be closing—through institutional guidance, emerging legislation and the first vacatur attempts based on alleged AI outsourcing. This article surveys that landscape and identifies the most pressing ethical concerns for arbitrators.

Several leading arbitral institutions and professional bodies have issued guidance specifically addressing AI use by arbitrators. The Silicon Valley Arbitration and Mediation Center (SVAMC) was among the first, releasing its Guidelines on the Use of Artificial Intelligence in Arbitration in April 2024. For arbitrators, the SVAMC Guidelines establish two foundational obligations: non-delegation of decision-making and respect for due process.

Several institutions have since issued AI guidance directed at arbitrators, including the American Arbitration Association (AAA)-International Centre for Dispute Resolution (ICDR), the Vienna International Arbitral Centre (VIAC) and the Stockholm Chamber of Commerce (SCC). The Chartered Institute of Arbitrators (CIArb) also published its Guideline on the Use of AI in Arbitration in 2025.

Binding legal frameworks are also emerging. California Senate Bill 574 would codify several ethical obligations from the SVAMC Guidelines, including prohibiting arbitrators from delegating any part of their decision-making to generative AI; prohibiting reliance on AI-generated information outside the record without disclosure to the parties; and requiring arbitrators to assume responsibility for all aspects of an award regardless of AI assistance. The bill also addresses hallucinations, providing that if an AI tool cannot cite independently verifiable sources, an arbitrator shall not assume such sources exist or are accurately characterized.

Recital 61 of the EU AI Act (Regulation (EU) 2024/1689), which entered into force in August 2024, applies to arbitrators using AI tools for “researching and interpreting facts and the law and in applying the law to a concrete set of facts,” as “AI tools can support the decision-making... but should not replace it: the final decision-making must remain a human-driven activity.” The Act excludes “purely ancillary administrative activities that do not affect the actual administration of justice.” Id. It requires tribunals using covered AI tools to take certain steps, including retaining automatically generated logs for at least six months and informing the parties about the use of high-risk systems. Violations can carry fines of up to €15 million.

Four core ethical obligations

Across these guidelines and legal frameworks, four ethical obligations emerge as most critical for arbitrators using AI.

1. Non-delegation of decision-making

The most consistently emphasized obligation is that arbitrators must not delegate their decision-making function to AI. It has long been recognized that arbitrators may not delegate their duties to a third party and must take personal responsibility for their awards. See, for example, 2004 AAA Code of Ethics for Arbitrators in Commercial Disputes, Canon V(C); 2017 SCC Rules, Art. 24(2); ICC, Note to Parties and Arbitral Tribunals on the Conduct of the Arbitration Under the ICC Rules of Arbitration ¶223 (2021); LCIA, Notes for Arbitrators ¶68 (2017) (“[I]n no circumstances may an Arbitral Tribunal delegate its fundamental decision-making function”). While guidance on this principle traditionally addressed human third parties such as tribunal secretaries (see, for example, ICC, Note to Parties and Arbitral Tribunals on the Conduct of the Arbitration Under the ICC Rules of Arbitration, ¶¶216-30 (2021)), that guidance naturally extends to AI tools.

SVAMC Guideline 6 states that “an arbitrator shall not delegate any part of their personal mandate to any AI tool,” and that AI tools “shall not replace their independent analysis of the facts, the law, and the evidence.” The AAA-ICDR similarly advises that “AI-generated outputs must never substitute for the arbitrator’s independent legal and factual analysis.” The CIArb Guidance more specifically warns that arbitrators should avoid “delegating any tasks to AI Tools, such as legal analysis, research and interpretation of facts and law, or application of the law to the facts, if such use could influence procedural or substantive decisions.”

The risk of inadvertent delegation is real. As the SVAMC cautions, arbitrators should review any AI output to ensure it reflects—and does not replace—their “independent analysis of the issues and evidence at hand.” The critical question is not whether AI is used, but whether its use crosses the line from assistance into substitution for the arbitrator’s own judgment. This principle is rooted in the foundation of the arbitrator’s mandate: parties agreed to submit their dispute to a specific human decision-maker, not to a machine.

That is, of course, unless the parties have specifically consented to an AI arbitrator, such as that recently introduced by the AAA. In this format—currently limited to lower value, document-only, opt-in construction cases—an AI arbitrator drafts a proposed award that is reviewed, finalized and issued by a human arbitrator. The parties may review the AI’s summary of their cases before it drafts a proposed award, but the human arbitrator ultimately remains responsible for the content of the final award.

2. Safeguarding confidentiality

The second major obligation relates to confidentiality—a cornerstone of arbitration that AI use can readily undermine. When an arbitrator submits case materials to a publicly available AI tool, those materials may be retained, reused for model training or potentially accessed by third parties.

SVAMC Guideline 2 provides that all participants “are responsible for ensuring their use of AI tools is consistent with their obligations to safeguard confidential information” and “should not submit confidential information to any AI tool without appropriate vetting and authorization.” CIArb’s Guidelines note that inputting confidential data into AI tools raises “concerns about how such data is stored, protected from outside access, used for machine learning purposes and potentially presented to third parties as output from their use of the same AI platform.”

The AAA-ICDR prohibits arbitrators from uploading “any confidential or identifying case materials into publicly available AI systems that do not provide contractual assurances against data reuse, model training, or third-party access.” The SCC Guidelines note that “the use of some AI tools may have unintended consequences for the confidentiality of an arbitration” and encourage arbitrators to understand how any data input is employed by the AI tool.

Arbitrators should understand the privacy and data retention policies of any AI tools they use and confirm they comply with the arbitrator’s confidentiality obligations. Arbitrators should only use closed or enterprise AI tools with contractual assurances against training on inputs and limited data retention (that is, not the free, standard version of ChatGPT); redact or anonymize case-specific information; and treat any public-facing AI tool as categorically unsuitable for identifiable case materials.

3. Transparency and disclosure

The third core obligation is transparency. AI hallucinations and bias are well-documented issues, and because AI use can affect the substantive reasoning underlying a decision, parties have a due-process interest in knowing how arbitrators have used AI and in having an opportunity to respond or comment on that usage.

SVAMC Guideline 7 provides that “an arbitrator shall not rely on AI-generated information outside the record without making appropriate disclosures to the parties beforehand and, as far as practical, allowing the parties to comment on it.” The VIAC Note provides that arbitrators may inform the parties how they plan to use AI tools, and the parties must have an opportunity to comment; it also suggests addressing AI use in Procedural Order No. 1. CIArb goes further, recommending that unless the parties agree otherwise, arbitrators should consult with the parties on AI tool use and provide an opportunity to comment on, and oppose, such use.

Available guidance typically does not require disclosure of administrative or clerical AI use, such as grammar checking or formatting. Rather, generative AI use that risks substantively influencing an arbitrator’s reasoning or findings approaches the line of required disclosure. Arbitrators must carefully monitor this line as AI tools become more integrated into their workflow.

4. Accuracy, verification, and the hallucination problem

Generative AI can produce incorrect citations, plausible but false factual statements and summary errors that subtly change meaning. Large language models may “hallucinate”—generating fluent, coherent responses based on probabilistic patterns without assessing accuracy. In an arbitral award, an undetected hallucination could be reputationally damaging or legally significant.

This risk is uniformly recognized across institutional guidance. The SCC Guide notes that arbitrators should “ensure effective human oversight” and review AI outputs to ensure it does not impact award quality. SVAMC Guideline 7 provides that arbitrators must not assume an AI tool’s cited sources actually exist or are accurately characterized.

Arbitrators should never adopt an AI-generated citation without personally reading the source, should verify every factual finding against record evidence, and should treat AI legal research as a starting point, not a conclusion.

The risk of award challenges: LaPaglia v. Valve Corporation

The abstract risk of an award challenge based on alleged AI use became concrete in LaPaglia v. Valve Corporation, filed in the US District Court for the Southern District of California in April 2025. See Petition to Vacate Arbitration Award, LaPaglia v. Valve Corporation, No. 3:25-cv-833 (S.D. Cal. 2025).

Petitioner LaPaglia sought to vacate a AAA award, alleging that the arbitrator outsourced his adjudicative role to AI, thereby exceeding his powers. Id. at 8. LaPaglia relied on the arbitrator’s alleged statements about using ChatGPT to save time writing articles, comments about wanting to finish the case before an upcoming vacation, and several factual statements in the award that were not argued by the parties and lacked record citations. Id. These passages were also run through ChatGPT, which allegedly identified them as AI-generated. Id. at 9. LaPaglia argued that the parties agreed to “a neutral arbitrator” rendering “a written decision” with a “statement of reasons,” and that reliance on AI betrayed this expectation and constituted an impermissible delegation of the arbitrator’s duties warranting vacatur. Id. The court ultimately dismissed the petition for lack of subject matter jurisdiction without reaching the merits.

Despite the procedural dismissal, LaPaglia illustrates how an arbitrator’s AI use could provide a basis for party to seek vacatur. Whether such a claim could succeed on the merits—or if such evidence would be sufficient to meet the high bar of vacatur—remains to be seen.

Conclusion

While the promise of increased efficiency, cost reduction and enhanced analytical capability is real, so too are the ethical risks associated with AI’s use. The institutional consensus is clear: AI can assist the arbitrator, but it cannot be the arbitrator without consent of all parties, and arbitrators must remain responsible for their awards. As guidance becomes codified into law and awards begin to face challenges on the basis of AI usage, arbitrators must remain vigilant in ensuring their use of AI tools complies with their ethical obligations.

Issue 24

  • content-type:insights/publication
  • nrf-migrated:article-categories/hubs/international-arbitration-report/issue-24
card-grid-item, aem-sourced
International arbitration report
/en/insights/publications/international-arbitration-report-hub

Contact