Blaise Pascal was rumored to have once written in a letter, “Je n’ai fait celle-ci plus longue que parce que je n’ai pas eu le loisir de la faire plus courte.” (I only made this one longer because I didn't have the time to make it shorter). If only he knew what was coming.

In 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, 2026 QCCS 1360, the Quebec Superior Court annulled an arbitral award that relied exclusively on AI-hallucinated and irrelevant authorities, finding it breached procedural fairness.

Background

The underlying dispute arises from the application of dispute resolution provisions in a framework agreement (the “National Agreement”) between the co-Applicant, the Association des ressources intermédiaires d'hébergement du Québec (“ARIHQ”), which represented intermediary resource providers, and the Respondent, Santé Québec - Centre intégré universitaire de santé et de services sociaux du Centre-Sud-de-l'Île-de-Montréal (“Santé Québec”).

The National Agreement sets out the methods and scales of remuneration for services provided to Santé Québec by the ARIHQ’s members. The co-Applicant, the Osman Health Centre (“Osman”), was an intermediary resource provider and one of the ARIHQ’s members.

The National Agreement also sets out a dispute resolution process culminating in arbitration. If the parties could not resolve a dispute amicably, the ARIHQ or its relevant member must submit a written notice of disagreement to Santé Québec within 90 days. Santé Québec must then respond within 30 days. If Santé Québec’s response did not resolve the dispute, ARIHQ had to submit the dispute to arbitration within 60 days.

Osman provided adult accommodation services to Santé Québec under two sub-agreements (the “Sub-Agreements”). It sent a request for retroactive payment for services rendered to Santé Québec for which it asserted it had not been paid (the “Payment Request”). Santé Québec refused the request. The parties met in September 2021 to attempt a resolution, but Santé Québec maintained its position.

ARIHQ sent a notice of disagreement to Santé Québec, citing its failure to pay Osman for services rendered under the Sub-Agreements. Santé Québec responded on August 28, 2024, refusing payment and stating that the deadline for submitting the notice of disagreement expired 90 days after Santé Québec provided its final refusal.

Osman and ARIHQ filed a notice of arbitration. In response, Santé Québec filed an application to dismiss, arguing that the notice of disagreement was filed out of time.

On August 8, 2025, the Arbitrator issued his award (the “Award”) dismissing Osman and ARIHQ’s claim, finding that the notice provisions of the National Agreement were valid. In the Award’s analysis section, it cited three non-existent court cases, a fake academic article, a non-existent arbitral award, and a labour tribunal decision that did not support the proposition for which it was cited.

The annulment application

Osman and ARIHQ challenged the Award before the Superior Court of Quebec on two grounds:

  1. That the Award was contrary to public policy and the Civil Code of Quebec’s (C.c.Q) prohibition against contractually altering prescriptive periods (limitation periods).
  2. That the decision did not respect arbitral procedure as a result of the arbitrator delegating part of his decision-making authority to artificial intelligence.

The Court dismissed the public policy ground, holding that an arbitrator’s error of law, even on a matter of public policy, was not a basis for annulment. The Court also found that, in any event, the 90-day notice period was a valid contractual claim procedure that did not modify the limitation period.

The Court nevertheless set the Award aside on the procedural fairness ground. It found that the Arbitrator’s reliance on AI to draft the Award, which was evident from non-existent and incorrectly cited authorities, violated procedural fairness. This was sufficient to warrant annulment under article 646(3) of the Code of Civil Procedure, as “the applicable arbitration procedure was not observed.”

The Court rooted its reasoning in the audi alteram partem principle. It held that decision-makers cannot delegate their authority, including the responsibility for writing reasons, whether to AI or otherwise. Specifically:

  • The Court held that delegating the drafting of the Award violated the parties’ right to choose their arbitrator. Parties select arbitrators for a variety of reasons, including legal expertise, decision-making experience, and availability. Parties are entitled to expect that the arbitrator they choose will hear and decide their dispute.
  • The Court emphasized the importance of written reasons, which ensure due consideration of the issues, reduce the risk of arbitrary decisions, and strengthen public confidence in judgments. Just as parties are entitled to expect their chosen arbitrator will hear and decide their dispute, they are likewise entitled to their arbitrator’s reasons.
  • The Court found that delegating the drafting of the Award violated deliberative secrecy. An arbitrator’s duty to maintain deliberative secrecy exists to ensure that their award reflects the arbitrator’s own opinion. The Court was careful to note that deliberative secrecy does not prevent an arbitrator from using researchers, clerical staff, translators, or similar means of assistance, nor does it prevent a decision-maker from using colleagues as a sounding board. The line is crossed, however, where someone or, in this case, something other than the arbitrator effectively steps in and takes over the reasoning and writing process.

Analysis

This case reinforces that the ultimate responsibility for reasoning and writing a decision rests with the arbitrator. Delegating that responsibility risks creating the appearance that a third party is ultimately in control of the decision-making process. Here, the fact that 100 per cent of the authorities cited in the Arbitrator’s analysis were either hallucinated or not on point demonstrated to the Court that the Arbitrator went beyond using AI as an assistive aid, and strayed into fully delegating the decision-making process.

The Court also noted that AI use will not automatically result in annulment in every case. This is so even where AI has been used as a drafting tool or has led to hallucinated authorities. It remains necessary to balance the breach of procedural fairness against its impact on the award. Where the breach could not reasonably have had any impact on the outcome, the court may exercise its discretion to recognize the procedural fairness defect, but maintain the award.

The Court also used this case as an opportunity to remind the legal profession of the risks associated with AI use, both by practitioners and decision-makers, beyond the commonly reported hallucinations. The Court noted:

  • The law often requires decision-makers to use their discretion, considering such factors as community values, the parties’ subjective characteristics, and other relevant contextual circumstances. AI outputs are created by programming and generally cannot account for human values that vary with the circumstances, particularly in novel situations. If decisions appear to have been made by AI without these considerations in mind, public confidence in the law may be shaken.
  • Sometimes biases are unintentionally encoded into AI tools by the material used to train the AI. Without access to the AI system’s source code, accounting for these biases may be impossible. This makes it all the more important that practitioners and decision-makers take responsibility for their arguments and analysis.
  • AI systems often lack confidentiality. Materials uploaded into an AI tool enter its database and may be disclosed to other users. Where that information is confidential, practitioners risk violating the confidentiality that underlies the solicitor-client relationship, while an arbitrator may breach the obligation to maintain deliberative secrecy. Indeed, a similar, though somewhat different issue recently arose in U.S. v. Heppner, in which the United States District Court for the Southern District of New York held that written exchanges an accused had with the Claude AI platform about his litigation were protected by neither solicitor-client privilege nor work product privilege (litigation privilege).

In summary, all actors in the justice system must take great care when relying on AI tools to develop, argue or decide a case. The Court’s decision emphatically demonstrates that arbitrators are no exception.

As courts continue to address the role of AI in legal proceedings, organizations and legal professionals should monitor these developments closely. To discuss the implications of this decision, please contact one of the authors or a member of Gowling WLG's Commercial Litigation and Arbitration Group.