Introduction
Nowadays, it is not surprising that almost every counsel preparing submissions, every expert producing a report, and every tribunal drafting an award relies to some degree on Artificial Intelligence (AI). Whether the benefits of AI in arbitration outweigh its risks remains genuinely contested (American Arbitration Association, 2026). However, given that in recent years major arbitral institutions have responded to this innovation by developing guidelines and setting limits on its use, the question of whether AI should be embraced has already taken a back seat to the question of how it should be constrained.
This growing institutional response, however, exists alongside a legal framework that was never designed with AI in mind. While arbitral institutions can adapt to emerging technology through internal guidance, one instrument cannot adapt so smoothly. This is the Convention on the Recognition and Enforcement of Arbitral Awards (NYC), which was adopted in 1958. At that time, the idea that a machine can assist or even generate an arbitral award would have belonged to science fiction (Gielen, 2026).
The tension between modern arbitral practice and this older enforcement framework has become particularly visible through two recent developments. Firstly, the AAA-ICDR’s introduction of its “AI Arbitrator” in November 2025. Secondly, the April 2026 decision of the Quebec Superior Court setting aside an award involving AI-generated reasoning.
This paper argues that these two developments rest on the same unresolved legal problem. It does not examine whether parties may consent to a fully autonomous AI arbitrator, nor every potential ground for non-enforcement under the NYC. The paper proceeds from the approach, now shared across multiple institutions, that AI assistance is permitted but that the appointed arbitrator must personally retain the decision-making function (Meredith et al., 2026). Against that background, it considers how to distinguish AI assistance from AI delegation in light of Article V(1)(d) of the NYC.
One Issue, Not Two Positions
At first glance, AAA-ICDR’s “AI Arbitrator” and the Quebec Superior Court’s annulment of an arbitral award seem to contradict each other. However, on closer examination, they point in the same direction. The “AI Arbitrator” has a narrow scope and is available only for two-party, documents-only construction disputes. It is not designed to fully automate the preparation of awards. Rather, it operates within a “human-in-the-loop” framework, in which a trained human arbitrator reviews, refines, and issues the final decision (Morgan et al., 2025).
ARIHQ v Santé Québec, decided in April 2026, points in the same direction. The court held that an arbitrator had delegated decision-making authority to AI and annulled the award on that basis (ARIHQ v. Santé Québec, 2026, para 114-116). Importantly, in that case, the award relied on legal authorities that AI had hallucinated, and those authorities were central to the reasoning and directly shaped the outcome (Huang et al., 2026). The court did not prohibit the use of AI as such. In contrast, it drew a firm line: AI may assist, but it must not drive, the decision-making process. Furthermore, the court was careful to add that not every award containing erroneous references, or drafted with some help from AI tools, would be annulled.
This distinction is not limited to one court decision. Other institutions have adopted a similar position. For example, the ICSID tribunal, in its procedural order, set out the tasks AI could permissibly assist with, while making clear that any delegation of actual decision-making to AI remains prohibited (ENCORE Investment Group Ltd. v. Republic of Türkiye, 2025, para 28.1). The Vienna International Arbitration Centre has issued a non-binding “Note on the Use of AI” that recognises its usefulness while stressing the continuing need for human oversight (VIAC, 2025, para 2.1). In addition, the American Arbitration Association’s Guidance on Arbitrators’ Use of AI Tools rests on the same principle of preserving the arbitrators’ independent decision-making (AAA-ICDR, 2025, para 3).
Taken together, these developments reflect a common approach in current arbitral practice: AI may assist, but independent human decision-making should remain intact. The unresolved question is, however, how to identify the line between permissible assistance and impermissible delegation, particularly once the award reaches the enforcement stage.
Navigating the AI Non-Enforcement Hurdle
Considering the scenario where the parties agreed that AI may assist but not replace the decision-making power of an arbitrator, or where the law of the seat prescribes the same (as it was in the ARIHQ v Santé Québec), Article V(1)(d) of the NYC may serve as a ground to refuse the recognition and enforcement of the AI-generated award. In that case, the question is no longer about “AI or human” but whether the process matched the promise the parties agreed to (NYC, 1958, art V(1)(d); Gielen, 2026). The main challenge is to determine the threshold that should be reached before AI involvement amounts to a procedural irregularity serious enough to justify non-enforcement in a State other than the seat of arbitration.
In determining that threshold, ARIHQ v Santé Québec may offer valuable guidance to NYC Contracting States, since the basis for the annulment of the award, Article 648 read with Article 646 of the Quebec Code of Civil Procedure, closely mirrors Article V(1)(d) of the NYC and Article 34(2)(a)(iv) of the UNCITRAL Model Law (Liu, 2026). While Article V(1)(d) of the NYC does not directly specify the threshold of procedural irregularities, most courts require a substantial defect in the arbitral proceeding, or a causal link between the defect and the award (Liu, 2026). Consequently, the NYC does not “permit reviewing courts to police every procedural ruling made by the arbitrator” (Compagnie des Bauxites de Guinée v. Hammermills, Inc., 1992).
Although existing practice, and ARIHQ v Santé Québec in particular, helps to define the threshold in legal terms, proving that the arbitrator delegated its decision-making power to AI tools may be considerably harder. Compared to ARIHQ v Santé Québec, subsequent cases may be less straightforward. What if an AI-assisted award contains correct authorities and plausible reasoning while still giving little indication of whether the arbitrator independently reviewed and adopted that reasoning?
Most institutions have not been silent on the underlying duties of arbitrators when AI tools form part of the decision-making process. CIArb’s 2025 Guideline on the Use of AI in Arbitration clearly addresses several considerations, recommending, for example, that arbitrators verify the accuracy of information obtained through AI, ensure that AI-assisted output is free from bias, and consult with the parties on its usage (CIArb, 2025 paras 8.3, 9.1). SVAMC’s guidelines likewise address disclosure, confidentiality, and due diligence (SVAMC, 2024). This soft law matters for more than institutional self-regulation. Where the seat of arbitration has no case law on arbitrators’ delegation to AI, a widely adopted institutional standard may serve as an important source that courts and tribunals can consider. What none of these instruments addresses, however, is how to establish whether the relevant duty was actually discharged.
Solution
Existing guidelines cannot fully police the line between AI assistance and delegation, because what they ask arbitrators to demonstrate is, in practice, extremely difficult to prove. “Independent” and “free from bias” describe states of mind, and a state of mind cannot be produced on request once an award is challenged. Only a record of an action can. Institutional guidance on AI use by arbitrators should therefore be drafted with Article V(1)(d) of the NYC in mind, since a duty phrased as a documented procedure gives an enforcing court something concrete to examine. The proposed three specific requirements, each replacing a duty currently expressed only as an adjective, are the following:
Independent verification – a dated record showing that each authority that AI supplied was checked against its primary source.
Traceable authorship – retention of the AI-generated draft and the arbitrator’s own edited version as separate, timestamped files, so that a later reviewer can see precisely what the arbitrator changed, added, or rejected. The records should remain securely retained and confidential (ICC, 2026, art 12.8), with access granted only following a sufficiently substantiated challenge to the arbitrator’s independent decision-making.
Task-specific disclosure – disclosure of the particular function AI performed at an agreed stage of the proceedings.
Conclusion
To conclude, the growing use of AI in arbitration does not necessarily threaten the enforceability of arbitral awards. The real difficulty arises when AI assistance becomes so extensive that it is no longer clear whether the arbitrator has retained the decision-making function. ARIHQ v Santé Québec demonstrated this problem in an obvious form, but future cases may involve legally coherent awards that reveal little about how much independent judgment the arbitrator actually exercised.
This uncertainty becomes particularly important at the enforcement stage. Article V(1)(d) of the NYC may provide a route to challenge such awards where the agreed arbitral procedure required the arbitrator to retain independent decision-making authority. However, that protection is only meaningful if the relevant human involvement can actually be demonstrated. Independent verification, traceable authorship, and task-specific disclosure would therefore turn the existing abstract duties of human oversight into something an enforcing court can assess in practice.
References
American Arbitration Association. (2026). Artificial intelligence in arbitration: Is there room for AI arbitrators? – Part I. https://www.adr.org/news-and-insights/ai-in-arbitration-is-there-room-for-ai-arbitrators/
Meredith, L., Thompson, N., & Greenaway, O. (2026, April 22). AI in arbitration: Institutional guidance and emerging developments. Society for Computers & Law. https://www.scl.org/ai-in-arbitration-institutional-guidance-and-emerging-developments/
Morgan, C., de Sousa, M., & Melville, K. (2025). The AAA-ICDR’s AI Arbitrator: A new chapter in dispute resolution? Herbert Smith Freehills Kramer. https://www.hsfkramer.com/notes/arbitration/2025-11/the-aaa-icdr-ai-arbitrator-a-new-chapter-in-dispute-resolution.
Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) c. Santé Québec – Centre intégré universitaire de santé et de services sociaux du Centre-Sud-de-l’Île-de-Montréal, 2026 QCCS 1360.
Huang, J., Abu-Manneh, R., Mata Morreo, G., Harrington, B. J., Morgan, R., & Dubot, L. (2026, May 13). Canadian court annuls arbitral award for delegation to AI, consistent with global trends. MayerBrown. https://www.mayerbrown.com/en/insights/publications/2026/05/canadian-court-annuls-arbitral-award-for-delegation-to-ai-consistent-with-global-trends.
ENCORE Investment Group Limited (Malta) v. Republic of Türkiye, ICSID Case No. ARB/24/46, Procedural Order No. 1 (July 22, 2025).
Convention on the Recognition and Enforcement of Foreign Arbitral Awards, June 10, 1958, 330 U.N.T.S. 3.
Vienna International Arbitral Centre. (2025). VIAC Note on the Use of Artificial Intelligence in Arbitration Proceedings. /api/wp-media/2025/04/VIAC-Note-on-AI-1.pdf
American Arbitration Association & International Centre for Dispute Resolution. (2025, March). Guidance on arbitrators’ use of AI tools.
Gielen, N. (2026, July 13). Debate: Are AI-assisted awards enforceable under the New York Convention? Kluwer Arbitration Blog. https://legalblogs.wolterskluwer.com/arbitration-blog/debate-are-ai-assisted-awards-enforceable-under-the-new-york-convention/
Liu, X. (2026, June 19). ARIHQ v Santé Québec: When must an AI-drafted award be set aside? Kluwer Arbitration Blog. https://legalblogs.wolterskluwer.com/arbitration-blog/arihq-v-sante-quebec-when-must-an-ai-drafted-award-be-set-aside/
Compagnie des Bauxites de Guinée v. Hammermills, Inc., No. 90-0169, 1992 WL 122712 (D.D.C. May 29, 1992).
Chartered Institute of Arbitrators. (2025, September). Guideline on the use of AI in arbitration. https://www.ciarb.org/media/bpndtcgu/guideline-on-the-use-of-ai-in-arbitration_updated-sept-2025.pdf
Silicon Valley Arbitration & Mediation Center. (2024). Guidelines on the use of artificial intelligence in arbitration (1st ed.). /api/wp-media/SVAMC-AI-Guidelines-First-Edition.pdf
International Chamber of Commerce. (2026). ICC Arbitration Rules, art. 12(8). https://iccwbo.org/dispute-resolution/dispute-resolution-services/arbitration/rules-procedure/2026-arbitration-rules/
