In Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) c. Santé Québec, 2026 QCCS 1360, the Superior Court of Québec annulled an arbitral award informed by non-existent legal authorities, likely generated by artificial intelligence. The Court considered this to have constituted a serious breach of the agreed upon arbitral procedure and the integrity of the decision-making process. The Court emphasized that, although arbitration awards are generally final and immune from review on the merits, courts may intervene where there is a fundamental procedural defect.
Facts
Osman operates intermediate housing resources. ARIHQ, on behalf of members such as Osman, and the Minister of Health and Social Services were parties to a framework agreement that contained a dispute resolution process with a 90-day limit to submit any notice of disagreement. Osman had in 2021 made a request for retroactive compensation to the Centre intégré universitaire de santé et de services sociaux du Centre-Sud de l’Île-de-Montréal (CCSMTL) (now Santé Québec), which the CCSMTL refused that year. In August 2024, ARIHQ then notified a notice of disagreement, followed by a notice of arbitration filed in September 2024, which resulted in an arbitration proceeding. On a preliminary basis, the arbitrator dismissed the claims as having been brought out of time.
Both Osman and ARIHQ then brought annulment proceedings in the Superior Court, which were consolidated. They contended, first, that the award violated public policy because it effectively enforced a contractual limitation period (90 days) that was shorter than the three-year prescription period prescribed by the Civil Code of Québec. Second, they argued that the arbitral procedure had not been respected because the award relied on non-existent legal authorities, suggesting that the arbitrator had improperly relied on artificial intelligence.
The Court’s analysis
(i) The contractual time limits
The Court held that, even if the arbitrator made a legal error, including one involving public policy, this did not justify annulment. Courts cannot review the merits of the arbitrator’s reasoning and may intervene only if the result itself contradicts fundamental principles of Québec law (paras. 48–50, 55). That was not the case here. The Court further found that the requirement to send a notice of disagreement within 90 days was not a forfeiture (“délai de déchéance”) (para. 58). Instead, it was a procedural precondition—a step that must be completed before a claim can be brought. As a result, this requirement did not shorten the three-year prescription period under article 2925 of the Civil Code. The arbitrator’s reliance on these deadlines was therefore not contrary to public policy.
(ii) The use of AI
While the use of AI is not, in itself, a statutory ground for annulment, the Court framed the issue as a failure to respect the arbitral procedure under article 646 of the Civil Code, emphasizing that the question was whether the integrity of the process agreed upon by the parties had been compromised through the use of AI in drafting the decision and reliance on cases that AI had generated. Article 646 must be interpreted broadly enough to capture serious procedural deviations, including those not expressly enumerated, provided they affect the integrity of the arbitral process (“l’intégrité du processus arbitral”) (para. 71).
The Court’s two-step test
The Court articulated a two-step approach to these situations. First, the court should determine whether there has been a violation of the agreed-upon arbitral procedure that impacts the integrity of the adjudicative function. Second, if such a violation is established, the court should weigh the nature of the breach in light of the arbitral procedure, its impact on the integrity of the process and its potential effect on the award itself (para. 72).
Applying the first step, the Court held that the arbitrator must personally exercise the adjudicative function and author the reasons supporting the award. The Court emphasized that party autonomy includes not only the choice to arbitrate, but also the choice of a specific decision-maker. The parties are entitled to have their dispute decided, and the reasons formulated, by that arbitrator (paras. 73–75).
The Court held that the arbitrator’s reliance on non-existent, hallucinated authorities supported an inference that the arbitrator had failed to exercise independent judgement or to verify the sources, calling into question whether the reasoning was truly that of the chosen decision-maker.
The risks associated with AI
In this context, the Court outlined several risks with AI, including the potential generation of false authorities (“hallucinations”), lack of human judgement, inherent biases, risks to confidentiality and the erosion of public confidence in the justice system (para. 90).
The Court emphasized that while AI can be a useful tool, it must not replace the arbitrator’s duty to verify sources and independently exercise judgement.
Application of the test to the facts
The Court then applied the two-step framework to the facts of the case. On the first step, the arbitrator relied extensively on authorities that did not exist, including several cases and an arbitral award that could not be located (paras. 106–112). These authorities formed the legal foundation of the arbitrator’s reasoning (para. 113), and were not peripheral.
The Court inferred, on a balance of probabilities, that the arbitrator had either relied on an external tool without adequate verification or otherwise had not fully engaged in the reasoning process, amounting to a delegation of the adjudicative function (paras. 114–115). This was a breach of the fundamental requirement that the arbitrator personally decide the dispute and assume responsibility for the reasoning.
On the second step, the Court evaluated the nature and impact of this breach and concluded that the violation was serious and directly undermined the integrity of the arbitral process. The use of non-existent authorities was not a technical defect; it went to the legitimacy of the reasoning itself. This type of defect was capable of eroding the parties’ confidence in the decision and in arbitration more broadly (para. 120).
Finally, because these fabricated authorities were central to the arbitrator’s analysis, the breach likely had a material impact on the outcome. A proper verification of sources could have led the arbitrator to reconsider or modify the reasoning and, potentially, the conclusion. Accordingly, the Court annulled the arbitral award and ordered the parties to appoint a new arbitrator within 60 days (para. 122).
The key takeaways
The decision is significant for arbitration practice. Although the Court reached the relatively unremarkable conclusion that contractually agreed procedural deadlines should be enforced, it ultimately held that an arbitral award cannot stand where its reasoning relies on fictitious authorities, likely generated by artificial intelligence. In that sense, it reaffirms the narrow scope of judicial intervention in arbitral awards while clarifying that procedural integrity remains a critical element. It demonstrates that courts will intervene where the adjudicative function is compromised by unacceptable practices such as reliance on fabricated sources.
The decision also establishes an important precedent regarding the permissible limits of AI use in arbitration, particularly by arbitrators, confirming that its use is acceptable only to the extent that it does not displace the arbitrator’s duty to personally exercise their deliberative function. The case is a caution both to arbitrators and to counsel to ensure that they are properly reviewing and supervising AI work product, and that arbitrators are using AI within the scope of their appointment.
Two issues not addressed in detail in the award are potential breaches of confidentiality as a result of the purported use of an AI tool (it was not clear what platform was used to generate these hallucinated cases or the terms of use), and how such potential breaches could be addressed without breaching principles of deliberative secrecy. While the Court discusses deliberative secrecy in the context of the arbitrator not delegating his decision, the issue here is how the parties could learn what data was used in what AI platform if such use occurred as part of the arbitrator’s deliberations. If a public or non-proprietary AI tool was used to generate the non-existent legal authorities, were the prompts or other information provided to the platform specific and confidential to the parties and their agreements and dispute? Could this question be answered without eliciting information that is arguably captured by the arbitrator’s deliberations? The risks inherent with the increasing frequency of use of AI in all aspects of arbitration underscore the need for parties to carefully consider whether to address arbitrator use of AI at the outset in a Procedural Order or otherwise. Guidance for how to do so can be found in the Chartered Institute of Arbitrators’ Guideline on the Use of AI in Arbitration (2025).
One final thought—what if the AI utilized here had not generated fictitious authorities and the arbitrator had based their award on that work product alone? We suspect no one would have ever discovered this. Would the award still be compliant with our laws? This issue will not go away, we strongly suspect, as AI becomes ever more reliable and hallucinations disappear. Stay tuned.
For more information about this topic, please reach out to the authors, Michael Schafler, Rachel Howie, Chloe Snider, Ioana Jurca and Ekin Cinar.
