Disputes Insights #8: Arbitral Award Set Aside in Quebec Over AI‑Generated Citations – What It Means for Arbitration Users

On 22 April 2026, the Superior Court of Quebec annulled an arbitral award because the arbitrator had relied on hallucinated case law and doctrine. The judgment is Association des ressources intermédiaires d'hébergement du Québec (ARIHQ) c. Santé Québec, 2026 QCCS 1360, available here.

It appears to be one of the first reported decisions to set aside an award on the basis that it was drafted with the help of generative AI. For arbitration users in Switzerland and elsewhere, the reasoning is worth reading carefully – partly for what it says, and partly for what it deliberately does not say.

What happened

The underlying dispute concerned a Quebec health-care contract. The arbitrator dismissed the claim on a preliminary objection, finding that the claimant had missed a contractual notice deadline. The award itself was short, but the legal reasoning rested on a handful of authorities. When the losing parties tried to verify them, they could not.

A doctrinal article cited as central to the reasoning did not exist. Three case citations led, when looked up via the citation number, to entirely unrelated decisions. A cited arbitral award turned out to be fictional – SOQUIJ, the Quebec legal-information authority, confirmed there was no such award. Every doctrinal and jurisprudential source the arbitrator relied on was either invented or misattributed. Evidently, the references were the product of AI hallucinations.

The losing parties applied for annulment on two grounds: that the award violated public policy, and that the arbitral procedure had not been respected because the arbitrator had effectively delegated his decision to a generative AI tool.

The public-policy argument failed

The Court rejected the public-policy ground. The argument was that the contractual notice deadline conflicted with article 2884 of the Civil Code of Quebec, which prohibits parties from setting their own limitation periods. The Court accepted that 2884 is a public-policy rule. It still refused to annul on that basis.

The reasoning here matters. Under Quebec law, as under most arbitration regimes, an error of law – even on a public-policy rule – is not a ground for annulment. What counts is whether the outcome of the award is incompatible with public policy, not whether the arbitrator got the legal analysis right. A contractual notice requirement as a precondition to suing has been upheld by the Quebec Court of Appeal in other contexts. The operative part of the award here did not cross the public-policy line, even if the reasoning was shaky.

This is a useful reminder that public-policy challenges to awards are narrow almost everywhere. Switzerland is no different: the Federal Supreme Court reviews awards under article 190 PILA on a similarly restrictive basis. Yet, it also accepts that blatant violations of procedural principles may constitute a violation of (procedural) public policy.

The procedural argument succeeded

The Court annulled the award on the second ground –that there was an important breach of the agreed arbitral procedure that affected the integrity of the proceedings. The reasoning is built around three propositions.

Party autonomy in choosing the arbitrator. The parties had selected their arbitrator from a closed list of ten names. Having made that choice, they were entitled to expect that the person they chose would actually decide the case and write the reasons. They had not chosen anyone else, and certainly not a language model.

The point of writing reasons. The Court returned to a basic idea: drafting reasons is not a formality. It is what forces the decision-maker to test an initial impression against the law. An award whose reasons were generated by a tool the arbitrator did not check is not a reasoned award in any meaningful sense.

The prohibition on delegation. Arbitrators are allowed to use research assistants, clerks, translation tools, and notes prepared by counsel. They are not allowed to hand over the decision itself, or the drafting of its reasons, to a third party. The principle "delegatus non potest delegare" applies. The Court was clear that an AI tool sits in the same category as a human third party for these purposes.

On the facts, the Court found that the fabricated citations were not incidental – they were the only legal authority cited in support of the result. The most plausible explanation for that pattern was that the arbitrator had relied on an AI tool and had not verified what it produced. That amounted to a delegation of his function. The award was annulled and the parties were ordered to choose a new arbitrator within 60 days.

No bright-line rule against AI

This is the part of the judgment that deserves the most attention. The Court was careful not to say that every award touched by AI must be annulled, or that every fabricated citation is fatal. Justice Sheehan made the point explicitly: AI use may be minimal, or it may concern a side issue. The analysis is contextual.

The test the Court applied has three parts:

1. the nature of the breach in light of the arbitral procedure that was engaged;

2. whether the breach affects the integrity of the proceedings; and

3. the actual impact of the breach on the award.

Minor or peripheral lapses will not meet that threshold. What tipped this case over the line was the scale of the breach. Every authority underpinning the operative part of the award was invented. The Court found that a party could reasonably think a proper verification of the citations would have led the arbitrator to revisit his position. The breach went to the integrity of the reasoning, and was capable of undermining trust not just in this award but in arbitration more broadly. That combination – material, central, and corrosive of confidence – was what justified annulment.

A Swiss perspective

The Quebec ruling has no binding force in Switzerland, but the underlying principles travel well.

Article 190 PILA allows a Swiss-seated international award to be set aside if, among other things, the tribunal was irregularly constituted, the right to be heard was violated, or the result is incompatible with public policy. Domestic awards are reviewable on similar but slightly broader grounds under article 393 of the Swiss Code of Civil Procedure.

The Federal Supreme Court has not yet had to deal with an AI-drafted award. But the principles it would draw on – the personal nature of the arbitral mandate, the prohibition on delegating decision-making, the requirement of genuine reasons, the secrecy of deliberations – are well established.

If those principles are taken seriously, an arbitrator who outsourced the operative reasoning to an AI tool without checking it would be exposed on at least three grounds: irregular constitution of the tribunal, because the actual decision-maker would no longer be the appointed arbitrator; and a violation of the right to be heard, because the parties would not, in any real sense, have been heard by the person deciding their case; and violation of procedural public policy.

The Quebec proportionality test would also resonate in Switzerland. A passing use of AI to clean up a draft is not the same as a wholesale delegation of the reasoning. Swiss courts would likely draw the same distinction.

Conclusion

The Quebec judgment is not a ruling against AI in arbitration. It is a ruling against an arbitrator who used AI without checking what it produced and let invented authorities carry the reasoning. The Court was careful to leave room for sensible AI use, while drawing a line at delegation. That is, in the end, the right place to draw it – and Swiss law, on the principles already in place, would draw it in roughly the same spot.