Assistance or Delegation? The Québec Decision on AI and the Setting Aside of Arbitral Awards

Until recently, decisions such as LaPaglia v. Valve Corporation signalled that courts were beginning to grapple with the role of artificial intelligence in arbitral decision-making. Yet, they stopped short of interfering with the award itself. 

On 22 April 2026, the Court Supérieure of Quebec, per Justice Martin F. Sheehan, took a step further by setting aside an arbitral award rendered on 8 August 2025, effectively turning a developing concern within the arbitral community into a tangible ground for judicial intervention. 

The point is not that AI has no place in arbitration. It plainly does. Few would seriously argue that arbitration can, or should, remain insulated from tools that are already reshaping legal practice. The concern, instead, is attribution: when does an award cease to be the arbitrator’s own decision?

Inside the Québec Decision: From Facts to Annulment

The dispute behind Association des ressources intermédiaires d’hébergement du Québec v. Santé Québec, 2026 QCCS 1360, was far removed from the world of technology. It arose in Québec’s health and social services sector. It concerned a claim for retroactive remuneration brought by ARIHQ and Centre de Santé Osman under a contractual framework governing intermediate housing resources. The arbitrator dismissed the claim at a preliminary stage, finding that the applicable contractual time limits had not been respected. 

The applicants sought to set aside the award on two grounds. The first was framed as a public order argument: the contractual time limit allegedly operated as an impermissible modification of Québec’s statutory limitation period. 

The second, which ultimately proved decisive, was procedural: the award relied on non-existent doctrine and case law, raising the concern that the arbitrator had used generative AI in a manner incompatible with the arbitral procedure chosen by the parties.

The Court’s reasoning is significant precisely because it did not treat the case as an opportunity to reopen the merits of the dispute. Justice Sheehan expressly recalled the limited nature of annulment proceedings: the correctness, fairness or even reasonableness of the award is not, in itself, a ground for setting it aside. 

Even a decision that is “surprising, or even legally fragile” remains beyond review unless one of the statutory grounds for annulment is established. The public order argument, therefore, failed. Even if the arbitrator had erred in his treatment of Québec limitation rules, that would have amounted to an error of law, not a basis for annulment.

What could not be treated as a mere legal error, however, was the award’s reliance on non-existent legal scholarship and case law. Those references did not simply weaken the arbitrator’s reasoning; they gave the appearance of a legal analysis that had never in fact taken place. In other words, the problem was not that the arbitrator may have reached the wrong conclusion, but that the reasons supporting that conclusion appeared to have been generated, or at least adopted, without the verification and independent judgment that arbitral decision-making requires.

The Court found that a serious breach of the arbitral procedure may justify annulment where it affects the integrity of the arbitral process. In that context, the identity of the decision-maker mattered. By agreeing to arbitration, the parties had not merely chosen a private forum; they had chosen a particular arbitrator. They were therefore entitled to expect that the reasons disposing of their dispute would be the arbitrator’s own. In Justice Sheehan’s words, the combined importance of party autonomy, reasoned decision-making, deliberative secrecy and public confidence justified the requirement that arbitral decisions be drafted by the arbitrator chosen by the parties, “without delegation to third parties.”

That conclusion, however, should not be mistaken for a blanket rejection of AI-assisted arbitral work. Justice Sheehan expressly recognised that technological tools capable of improving access to justice should be “welcomed and regulated rather than prohibited and stigmatised”. The judgment also acknowledged that legal professionals increasingly use large language models to summarise documents, identify relevant materials, transcribe audio, refine drafting and assist with legal research.

The Court’s concern was therefore not the use of AI as such, but the point at which assistance becomes substitution.

The distinction becomes sharper when AI is used not by counsel, but by the decision-maker. A lawyer who relies on AI-generated material without verification may breach professional duties; an arbitrator who does so in the reasons for an award risks something more fundamental, the loss of control over the adjudicative function itself. That is why Justice Sheehan observed that the pitfalls are “even greater” for decision-makers.

On the facts, that line had been crossed. The fabricated authorities were not marginal. They were the only doctrinal and judicial references supporting the arbitrator’s conclusion on the validity of the contractual time limits. The Court therefore found, on the balance of probabilities, that the arbitrator had delegated his authority and had abdicated his role of reviewing the result. In that sense, the problem was not technological novelty, but loss of adjudicative control. The hallucinated authorities were not cosmetic defects; they went to the core of the reasoning. For the Court, that was a serious breach of the arbitral procedure, one that likely affected the outcome and undermined confidence in the process. The award was set aside, and the parties were sent back to arbitration before a new arbitrator.

Why It Matters: Trust and Finality

The significance of the judgment lies less in the fate of this particular award than in the questions it leaves for arbitration more generally. Two consequences stand out.

The first is trust: can parties remain confident that the arbitrator they selected actually decided the dispute? The second is finality: will AI-related defects become a new route for challenging awards that arbitration, by design, seeks to make resistant to review?

Arbitration depends on a particular form of trust. Parties accept a limited right of recourse because they have chosen the decision-maker. Once AI enters the deliberative space invisibly, that confidence becomes harder to sustain. The issue is not whether arbitrators may use technological tools; they already do, and AI has already accelerated that trend.

The issue is whether the parties can still trace the award back to the arbitrator’s own judgment.

This is why the Québec decision matters beyond its facts. It gives disappointed parties a new line of attack: not simply that the arbitrator was wrong, but that the arbitrator may not have truly decided. Non-existent citations are the clearest example, but the concern is wider. Generic reasoning, unexplained authorities, or a failure to verify AI-generated material may all fuel suspicion that the award is less an exercise of adjudication than an output adopted after the fact.

The second consequence is equally important. Arbitration derives much of its practical value from finality. A losing party is not meant to relitigate the merits by recasting disagreement with the reasoning as a procedural defect. Yet improper AI use may open precisely that door. Where a challenge is framed not as “the arbitrator was wrong”, but as “the award is not truly the arbitrator’s reasoning”, courts may be drawn into questions that sit uncomfortably close to the merits while formally remaining procedural. If non-existent authorities, unverified AI-generated analysis, or undisclosed reliance on generative tools can cast doubt on whether the award is truly the arbitrator’s own reasoning, AI-related objections may become an increasingly attractive basis for set-aside applications.

This is not a reason to resist AI. Quite the opposite. It is a reason for arbitral actors to use it better. The technological tools now available to lawyers and arbitrators are extraordinary: they can reduce costs, accelerate review, improve access to information and make proceedings more efficient. But their value depends on disciplined use. If AI is treated as an invisible shortcut rather than a controlled instrument, it may generate the very inefficiencies arbitration seeks to avoid: satellite litigation, mistrust and challenges to awards.

Responsible Use, Not Prohibition

The lesson of the Québec decision is therefore not prohibition, but responsibility. AI should be embraced where it assists; it should be controlled where it influences reasoning; and it should never obscure the basic premise of arbitration, that the award is the decision of the arbitrator chosen by the parties.

That responsibility must be practical, not merely rhetorical. If AI is to be used safely in arbitration, the relevant question is not only whether it may be used, but how, with which tools, and subject to what safeguards. Two safeguards matter most here: verification and control.

First, verification. Any authority cited in an award should be checked against reliable legal databases. This is not an AI-specific duty; it is part of legal reasoning itself. What AI changes is the scale of the risk: hallucinated authorities may look plausible enough to pass unnoticed unless verification is treated as non-negotiable.

Verification also begins with the choice of the tool. There is a meaningful difference between using a general-purpose chatbot and using a legal AI tool connected to reliable legal databases. The latter undoubtedly reduces the risk of hallucinated authorities, but even in this case, the arbitrator or counsel remains responsible for checking the existence, relevance and accuracy of any authority relied upon.

Second, control. AI may assist with summaries, translation, document organisation, drafting refinement or legal research. But it should not supply reasoning that is adopted without independent review. The arbitrator must remain able to explain, defend and own the reasoning. Put simply, the arbitrator may be assisted by a tool, but cannot become the reviewer of a tool’s decision.

These safeguards would not eliminate every risk. Nor would they prevent parties from raising opportunistic challenges. But they would make it harder to argue that the arbitrator lost control of the decision-making process. That is what matters. The future of AI in arbitration should not be built on prohibition.

As Justice Sheehan recalled, quoting Justice Morin, technological measures that facilitate access to justice should be “welcomed and regulated rather than prohibited and stigmatized.” The challenge, then, is one of attribution, verification and control: ensuring that, when an award is rendered, it remains unmistakably the arbitrator’s award.