I. Introduction
Artificial Intelligence (“AI”) tools have transformed the way lawyers work, significantly increasing efficiency and reducing the time spent on tasks. Unsurprisingly, the trend has also made its way into the field of arbitration. AI now assists arbitral tribunals in calculating procedural timetables, conducting legal research and summarizing complex and voluminous case files, among other tasks. It has even been discussed as a candidate for the arbitrator’s role itself.[1]
At the same time, the use of AI by arbitrators has raised serious questions and risks, a prominent one being the delegation of decision-making to AI tools in violation of the arbitrator’s personal mandate. This is not new. In fact, the question of whether a secretary’s drafting assistance essentially replaced the personal judgment of the appointed arbitrators and violated the parties’ right to select their decision-maker has been a topic of debate.[2] With the emergence of generative AI, it has taken on a new dimension.
This article examines the governing framework and an example of how the principle operates in practice.
II. The Soft-Law Consensus: Non-Delegation
The field has been shaped by a growing body of soft law instruments, particularly the guidelines established by arbitral institutions.
Across these instruments, the principle of non-delegation emerges as the clearest point of convergence, even though each guideline frames it slightly differently.
- According to the Chartered Institute of Arbitrators Guideline, the use of AI may create risks for the administration of justice, and arbitrators therefore “should not relinquish their decision-making powers to AI.”[3] The guideline also makes it explicit that arbitrators must bear responsibility for their decisions, particularly in light of the requirements of impartiality and independence.[4]
- The American Arbitration Association and International Centre for Dispute Resolution Guidance on Arbitrators’ Use of AI Tools encourages arbitrators to make use of AI technology, while requiring that they must not delegate their reasoned judgment and must retain authority over their own evaluation.[5]
- The Vienna International Arbitral Centre Note on the Use of Artificial Intelligence in Arbitration Proceedings frames the same concern in terms of retained control, requiring arbitrators to maintain full authority over the decision-making process.[6]
- The SCC Arbitration Institute Guide to the Use of Artificial Intelligence in Cases Administered Under the SCC Rules notes that AI tools cannot replace the decision-making process and prohibits the delegation of the decision to them.[7]
- The Silicon Valley Arbitration & Mediation Center Guidelines on the Use of Artificial Intelligence in Arbitration take an equally categorical position: an arbitrator “shall not delegate any part of their personal mandate to any AI tool.”[8]
In sum, the underlying rule is consistent: although AI may assist in research, arbitrators must exercise their decision-making powers by themselves.
III. The Hard-Law Foundation
While no hard law instrument directly addresses an arbitrator’s use of AI, the principle of non-delegation is firmly rooted in hard law.
Under Article V(1)(d) of the Convention on the Recognition and Enforcement of Foreign Arbitral Awards (“New York Convention”), recognition and enforcement of an award may be refused if the “composition of the arbitral authority or the arbitral procedure was not in accordance with the agreement of the parties.”[9] Article 34(2)(a)(iv) of the UNCITRAL Model Law on International Commercial Arbitration (“UNCITRAL Model Law”) and Article 15/A(1)(b) and (f) of the Turkish International Arbitration Law No. 4686 recognize a corresponding ground for setting aside an arbitral award.[10]
Accordingly, where the legal reasoning in the essential parts of the award is produced by AI rather than by the arbitrators themselves, the procedure is likely to be considered inconsistent with the parties’ agreement and the award open to challenge.[11] A related but different question arises where an AI is appointed as arbitrator in its own right. In that case the objection would concern the composition of the tribunal, although it has been argued that no such objection should succeed where the arbitration agreement expressly provides for an AI-arbitrator.[12]
A second potential ground is public policy. Under Article V(2)(b) of the New York Convention, recognition may be refused where enforcement would be contrary to the public policy of the enforcing State.[13] Article 34(2)(b)(ii) of the UNCITRAL Model Law and Article 15/A(2)(b) of the Turkish International Arbitration Law No. 4686 provide corresponding grounds for annulment.[14]
For instance, as AI depends on the data on which it is trained, outputs shaped by biased data may result in a discriminatory award, undermining the arbitrator’s impartiality and exposing the award to annulment or refusal of enforcement on public policy grounds.[15] Similarly, where AI is used to such an extent that it essentially assumes the arbitrator’s decision-making function, the resulting award may be contrary to public policy in jurisdictions requiring arbitrators to be natural persons.[16]
IV. The Principle in Practice
A. Lessons from the Tribunal Secretary Debate
Although arbitrators’ delegation of decision-making powers to an AI tool has scarcely been tested, the question has long been discussed in relation to the delegation of decision-making powers to tribunal secretaries.[17] A prominent example is the annulment proceedings in the Yukos case before the Dutch courts, where the Russian Federation asserted that the tribunal had delegated its powers to the tribunal’s assistant.[18] The Hague Court of Appeal rejected the challenge. While assuming that the assistant had made significant contributions to the drafting of parts of the awards, it held that his participation in the decision-making process itself had not been established.[19] According to the court, absent any agreement between the parties, it fell within the tribunal’s discretion to decide the extent to which it relied on an assistant in drafting the award, provided that the substantive decisions remained its own.[20] Courts in France, Germany, England, Switzerland and Belgium have taken a similar view, consistently dismissing claims that a secretary’s involvement infringed the parties’ due process rights.[21] Nonetheless, while the precise scope of a tribunal secretary’s permissible functions remains subject to some controversy, it is generally settled that arbitrators may not delegate their essential powers to a secretary or assistant, including the determination of the dispute.[22] The same principle applies where a tribunal relies on AI in a manner that effectively transfers its adjudicative responsibilities to the system. Where an award is drafted by an AI system without adequate human consideration of the facts or the law, it may be open to challenge on the ground that the tribunal was not properly constituted and may also amount to a violation of public policy.[23]
In this context, a recent Canadian decision offers a judicial assessment of an arbitrator’s delegation of decision-making powers to an AI tool.
B. ARIHQ v. Santé Québec
Decided by the Québec Superior Court on 22 April 2026, ARIHQ v. Santé Québec concerned an application to annul an arbitral award rendered in a dispute over unpaid remuneration under a healthcare services agreement.[24] The Arbitrator had dismissed the claim as out of time, holding that the Claimants had failed to comply with a contractual time limit in the Parties’ dispute resolution procedure.[25] The Claimants sought annulment on two grounds.[26] First, they asserted that the award was contrary to public policy as it imposed a contractual time limit shorter than the prescription period provided by law.[27] Second, they argued that the award relied on non-existent legal doctrine and case law, which indicated that it had been drafted using an AI tool, and that the Arbitrator had thereby delegated part of his decision-making authority in breach of the agreed arbitration procedure.[28]
In its ruling, the Court rejected the public policy argument, holding that annulment review cannot extend to the correctness of an arbitrator’s legal reasoning, but only to whether the outcome itself is irreconcilable with fundamental principles of public policy.[29] The Court added that, in any event, the contractual time limit the Arbitrator had applied did not conflict with the statutory prescription regime.[30]
By contrast, the argument on the arbitration procedure succeeded.
While the Claimants presented strong evidence that the award had been drafted using AI, the Respondent argued that the use of AI is not among the grounds for annulment listed in Article 646 of the Québec Code of Civil Procedure.[31] The Court accepted, in principle, that this provision does not expressly regulate the use of AI,[32] but held that this did not resolve the question, since Article 646(3) permits annulment where “the procedure for the appointment of an arbitrator or the applicable arbitration procedure was not observed.”[33]
The Court grounded the prohibition on delegating decision-making authority in three converging principles: the parties’ autonomy in selecting the arbitrator, the duty to give reasons, and the arbitrator’s duty to maintain the confidentiality of the deliberations.[34] In the specific case of AI, the Court noted that the use of AI brings with it certain risks, such as the creation of false citations, the absence of discretion and human values, bias, the loss of confidentiality and the erosion of public trust in the justice system.[35] The Court further recalled that courts have sanctioned the use of AI by parties to draft pleadings, treating it as a serious breach that undermines the integrity of the judicial process.[36] Specifically, the Court referred to its earlier notice, recalling that AI must be used with caution, that parties should rely exclusively on sources from court websites or commonly cited publishers, and that they must retain rigorous human oversight of their submissions.[37] The Court considered the risks to be even greater for decision-makers, since judicial opinions must convey the full authority of the decision-maker, and the delegation of the drafting of a judicial decision is expressly prohibited by the Canadian Judicial Council.[38]
The Court was careful to note that the prohibition does not extend to research assistance, clerical support, translation or citation checking, nor does it prevent a decision-maker from consulting with colleagues.[39] In the Court’s view, what matters is that the involvement of third parties should not compromise the integrity of the process and that responsibility for the reasons remains with the decision-maker.[40]
On the facts, the Court found that the Arbitrator’s key reasoning was shaped by articles and case law that were hallucinated by AI, concluding that the Arbitrator had delegated his authority and abdicated his role in reviewing the outcome.[41] The Court emphasized that its decision does not mean that every award citing erroneous references or benefiting from AI must be set aside; rather, the nature of the breach and the extent to which the integrity of the proceedings was compromised must be assessed.[42] In the present case, however, the procedural violation was significant, as a thorough review of the non-existent decisions, which were central to the award, could have prompted the Arbitrator to reconsider his position.[43] Accordingly, the Court annulled the award and ordered the Parties to select a new arbitrator within sixty days.[44]
V. Conclusion
There is a clear consensus among soft law instruments that the decision-making powers of arbitrators must not be delegated to AI. Hard law instruments do not address the question directly, but public policy and procedural irregularity grounds are broad enough to accommodate it. Where AI takes over the decision-making responsibility of the arbitrator, the award may be at risk of annulment for departing from the agreed composition of the arbitral tribunal or the agreed arbitral procedure, or on public policy grounds.
The recent ARIHQ decision is a clear illustration that this risk is not merely theoretical. The decision does not, however, prohibit the use of AI outright. According to the Court, whether an award produced with the assistance of AI will be set aside depends on the nature of the breach and its effect on the integrity of the proceedings.
Finally, the prohibition is not as wide as it seems. Carrying out tasks such as research, translation and citation checking with the help of AI remains permissible. It may even be desirable, since it can reduce the cost and duration of the proceedings.[45] In contrast, the arbitrator’s own assessment of the facts and the law cannot be delegated. Such delegation is, however, subject to proof. ARIHQ was, in this respect, an unusually easy case as the hallucinated authorities were both the breach and the evidence of it. In most cases, a party that suspects improper use of AI will not know what tools or data sources were used, or what instructions were given.[46] Therefore, where an arbitrator delegates their decision-making powers to AI without leaving such traces, a party may find the breach considerably harder to establish, and the Court’s test correspondingly harder to apply.
—–
[1] See Gizem Halis Kasap, “Can Artificial Intelligence (“AI”) Replace Human Arbitrators? Technological Concerns and Legal Implications”, Journal of Dispute Resolution, Vol. 2021, No. 2, 2021, pp. 237-240.
[2] Gary B. Born, International Commercial Arbitration, 3rd ed., Kluwer Law International, 2021 (updated February 2024) (“Born”), §13.07[B]; Anne-Catherine Hahn, “Limits to the Use of AI Tools by Swiss Arbitral Tribunals: A Tentative Analysis”, ASA Bulletin, Vol. 43, No. 4, 2025 (“Hahn”), p. 754.
[3] Chartered Institute of Arbitrators, “Guideline on the Use of AI in Arbitration”, CIArb, September 2025 (“CIArb Guideline”), https://www.ciarb.org/media/bpndtcgu/guideline-on-the-use-of-ai-in-arbitration_updated-sept-2025.pdf?utm_source=AI+Guidelines+Launch&utm_medium=News+Page, accessed 26 August 2026, p. 16, para. 8.2.
[4] CIArb Guideline, pp. 5-6, para. 2.4.
[5] American Arbitration Association and International Centre for Dispute Resolution, “AAA-ICDR Guidance on Arbitrators’ Use of AI Tools”, 2026 (“AAA-ICDR Guidance”), https://www.adr.org/media/y4hmsz2m/2026_aaa-icdr_guidance_on_arbitrators_use_of_ai_tools.pdf, accessed 26 August 2026, pp. 1-2.
[6] Vienna International Arbitral Centre, “VIAC Note on the Use of Artificial Intelligence in Arbitration Proceedings”, April 2025, https://www.viac.eu/wp-content/uploads/2025/04/VIAC-Note-on-AI-1.pdf, accessed 26 August 2026, p. 3, para. 2.
[7] SCC Arbitration Institute, “Guide to the Use of Artificial Intelligence in Cases Administered Under the SCC Rules”, 2024, https://sccarbitrationinstitute.se/wp-content/uploads/2024/12/scc_guide_to_the_use_of_artificial_intelligence_in_cases_administered_under_the_scc_rules-1.pdf, accessed 26 August 2026, p. 3.
[8] Silicon Valley Arbitration & Mediation Center, “Guidelines on the Use of Artificial Intelligence in Arbitration”, 1st ed., 2024, https://svamc.org/wp-content/uploads/SVAMC-AI-Guidelines-First-Edition.pdf, accessed 26 August 2026, p. 12, Guideline 6.
[9] Convention on the Recognition and Enforcement of Foreign Arbitral Awards, New York, 10 June 1958, 330 UNTS 3 (“New York Convention”), art. V(1)(d).
[10] UNCITRAL Model Law on International Commercial Arbitration 1985, with amendments as adopted in 2006 (“UNCITRAL Model Law”), art. 34(2)(a)(iv); Law on International Arbitration no. 4686 dated 21 June 2001 (Milletlerarası Tahkim Kanunu) (“Turkish International Arbitration Law”), art. 15/A(1)(b) and (f).
[11] Dominik Stefer and Victoria Fricke, “From algorithms to awards: exploring the technological and legal boundaries of AI’s contributions to the work of arbitrators”, Arbitration International, Vol. 41, No. 1, 2025 (“Stefer and Fricke”), pp. 66-67.
[12] Gülüm Bayraktaroğlu-Özçelik and Ş. Barış Özçelik, “Use of AI-Based Technologies in International Commercial Arbitration”, European Journal of Law and Technology, Vol. 12, No. 1, 2021 (“Bayraktaroğlu-Özçelik and Özçelik”), p. 11.
[13] New York Convention, art. V(2)(b).
[14] UNCITRAL Model Law, art. 34(2)(b)(ii); Turkish International Arbitration Law, art. 15/A(2)(b).
[15] Bayraktaroğlu-Özçelik and Özçelik, p. 13.
[16] Id.; Stefer and Fricke, p. 67; see also João Ilhão Moreira and Jiawei Zhang, “ChatGPT as a fourth arbitrator? The ethics and risks of using large language models in arbitration”, Arbitration International, Vol. 41, No. 1, 2025 (“Moreira and Zhang”), pp. 79-80.
[17] Born, §13.07[B]; Hahn, p. 754.
[18] Born, §13.07[B]; Yukos Universal Limited (Isle of Man) v. The Russian Federation (“Yukos”), PCA Case No. 2005-04/AA227, Judgment of The Hague Court of Appeal dated 18 February 2020, para. 6.6.1.
[19] Born, §13.07[B]; Yukos, para. 6.6.6.
[20] Yukos, para. 6.6.14.
[21] Born, §13.07[B].
[22] Id.
[23] Hahn, p. 754; Stefer and Fricke, pp. 66-67.
[24] Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) v. Santé Québec – Centre intégré universitaire de santé et de services sociaux du Centre-Sud-de-l’Île-de-Montréal (“ARIHQ v. Santé Québec”), Superior Court of Québec (Commercial Division), Case Nos. 500-11-066427-259 and 500-11-066498-250, 2026 QCCS 1360, Judgment dated 22 April 2026, paras. 19-30.
[25] Id., para. 29.
[26] Id., para. 3.
[27] Id., paras. 3.1, 41.1.
[28] Id., paras. 3.2, 41.2.
[29] Id., paras. 55, 64-66.
[30] Id., paras. 56, 61-62, 65.
[31] Id., paras. 68-69.
[32] Code of Civil Procedure (Québec), CQLR c. C-25.01, art. 646, which provides as follows:
The court cannot refuse to homologate an arbitration award or a provisional or safeguard measure unless it is proved that
(1) one of the parties did not have the capacity to enter into the arbitration agreement;
(2) the arbitration agreement is invalid under the law chosen by the parties or, failing any indication in that regard, under Québec law;
(3) the procedure for the appointment of an arbitrator or the applicable arbitration procedure was not observed;
(4) the party against which the award or measure is invoked was not given proper notice of the appointment of an arbitrator or of the arbitration proceedings, or it was for another reason impossible for that party to present its case; or
(5) the award pertains to a dispute not referred to in or covered by the arbitration agreement, or contains a conclusion on matters beyond the scope of the agreement, in which case only the irregular provision is not homologated if it can be dissociated from the rest.
The court cannot refuse to homologate the arbitration award on its own initiative unless it notes that the subject matter of the dispute is not one that may be settled by arbitration in Québec or that the award or measure is contrary to public order.
[33] ARIHQ v. Santé Québec, para. 70.
[34] Id., paras. 73-87.
[35] Id., para. 90.
[36] Id., para. 91.
[37] Id., para. 93.
[38] Id., paras. 95-97.
[39] Id., para. 87.
[40] Id.
[41] Id., paras. 103-115.
[42] Id., paras. 117-118.
[43] Id., paras. 120-121.
[44] Id., paras. 122-123.
[45] AAA-ICDR Guidance, pp. 1-2; Moreira and Zhang, pp. 73-74.
[46] Hahn, p. 756.
