You Couldn’t Make it Up: 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

I have been increasingly bullish about the use of artificial intelligence in public administration. In a piece on “Artificial Intelligence and Administrative Tribunals“, I suggested that generative AI, such as Chat GPT or Claude, could be used to enhance the justification of decisions by providing reasons that are even more justified, transparent and intelligible than a human decision-maker would be capable of providing if left entirely to their own devices.

Yet, I warned, ” there is always a risk that a lazy decision-maker will use generative AI to actually make decisions”. Hence my interest in a case where this appears to have happened, 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. This was a private arbitration rather than a statutory decision-making process, but it functions as a cautionary tale for anyone exercising adjudicative functions.

Here, the arbitral decision was quashed, because significant portions were generated by artificial intelligence, rather than by the arbitrator himself. This is a basic sub-delegation issue — the arbitrator was appointed as the decision-maker, not Chat GPT, so by relying on generative AI, the arbitrator did not exercise the function he was supposed to exercise. As the superior court (Sheehan J) put it, with some understatement: “Il va de soi qu’en choisissant l’arbitre qui va entendre leur différend, les parties sont en droit de s’attendre à ce que ce soit cet arbitre qui rende la décision” (at para. 75).

The superior court recognized that resort to aid in drafting reasons, from clerks and colleagues, is entirely appropriate but subject to the primordial condition that the assigned decision-maker actually makes the decision (at para. 87). Here, however, just reading the decision on its face revealed serious issues, as the arbitrator’s reasons referenced non-existent scholarly articles and judicial decisions (at paras. 102-112). This is a common problem with large language models (see also another cautionary tale here). Worse, these articles and decisions were the only articles and decisions relied on by the arbitrator, and went to the very heart of his legal analysis (at paras. 113-116). Accordingly, the decision had to be quashed.

In other cases, the superior court recognized, a decision might not necessarily have to be quashed for reliance on generative AI, for instance if any errors are peripheral (at para. 117). This nonetheless functions as a cautionary tale.

Moreover, the superior court’s analysis provides a roadmap for reviewing decisions that might be based on generative AI — where there are hallucinations on the face of a decision, or other indices of the use of generative AI, this may justify piercing the veil of deliberative secrecy and require the decision-maker to furnish an explanation of how generative AI was used, including the prompts entered into the large language model, with a view to determining whether the decision-maker actually made the decision at all. Difficult questions might also arise if the decision-maker incorporates into their decision submissions of the parties (as they are entitled to do: Cojocaru v. British Columbia Women’s Hospital and Health Centre, 2013 SCC 30, [2013] 2 SCR 357) that are based on or incorporate hallucinations or errors attributable to generative AI.

For administrative tribunals wishing to avoid such a result, there are various options, including prohibiting generative AI (a prohibition that could be implemented by requiring decision-makers to produce decisions on computers that cannot access generative AI) or using internal generative AI systems that deploy watermarks or other devices to indicate when AI has been relied upon to generate text. Some consideration should also be given to how to ensure that errors do not creep in as a result of decision-makers copying and pasting submissions based on generative AI. Clearly, the rise of generative AI means that administrative tribunals should be considering proactive steps to ensure that this cautionary tale is not repeated in public administration.

PS, I asked Claude to comment on my draft: “Tight and purposeful. The tone is appropriately dry without being arid, which suits the subject. Nothing that needs significant reworking”.

This content has been updated on May 13, 2026 at 16:14.