The Québec Superior Court has set aside an arbitral award after finding that the authorities relied upon by the arbitrator were generated by AI and did not exist. Every case and doctrinal source cited in the award was a hallucination, leading the court to conclude that the arbitrator's reasoning may have been different had the authorities been properly verified.
The decision is one of the clearest judicial warnings to date about the risks of uncritical AI use in arbitration. While AI has the potential to improve efficiency, it cannot replace the arbitrator's personal responsibility to evaluate the evidence, apply the law and produce a reasoned award.
The Québec decision
In ARIHQ v Santé Québec, the applicant argued that the arbitral process had been fundamentally compromised because the award appeared to have been drafted using AI. The court agreed, finding a serious procedural defect that undermined the integrity of the arbitration.
Although courts are generally reluctant to interfere with arbitral awards, the judgment demonstrates that judicial deference has limits. The finality of arbitration depends upon confidence that the tribunal has exercised its own judgment. Where there is evidence that the decision-making process itself has been compromised, the courts will intervene.
Independent judgment cannot be delegated
The decision rests on a fundamental principle of arbitration: arbitrators are appointed for their expertise, independence and judgment. Those responsibilities are personal and cannot be delegated to a third party, whether a junior lawyer, research assistant or AI system.
AI may be used to organise documents, summarise evidence or improve drafting efficiency. However, the arbitrator must independently assess the facts, verify legal authorities and determine the outcome. The legitimacy of arbitration depends on parties receiving a decision reached by the tribunal they selected—not by an algorithm.
This approach is reflected in the 2025 CIArb Guidelines on the Use of AI in Arbitration, which make clear that arbitrators remain fully responsible for both the reasoning and outcome of their awards, regardless of any AI assistance.
Enforcement implications
The decision also has potentially significant consequences for the international enforceability of arbitral awards.
Article V(1)(d) of the New York Convention permits recognition and enforcement to be refused where the arbitral procedure was not conducted in accordance with the parties' agreement or, failing such agreement, the law of the seat. A similar ground appears in Article 34(2)(a)(iv) of the UNCITRAL Model Law.
The Québec judgment suggests that the use of AI may move beyond a question of drafting quality and become a procedural issue where it affects the tribunal's exercise of independent judgment. If an arbitrator relies on fabricated authorities without verification, it raises the question whether the tribunal genuinely exercised the adjudicative function entrusted to it.
That distinction is important. An award containing an incorrect citation is not unusual and will rarely justify judicial intervention. By contrast, an award founded on non-existent authorities may indicate that the tribunal failed to conduct the legal analysis expected of an independent decision-maker. In those circumstances, the defect lies not simply in the accuracy of the reasoning, but in the integrity of the arbitral process itself.
The case therefore provides parties resisting enforcement with a potential basis for arguing that the tribunal failed to perform its mandate in accordance with applicable procedural standards.
The position under English law
Comparable issues could arise under English law.
Section 68 of the Arbitration Act 1996 permits an award to be challenged for a serious irregularity causing substantial injustice. The provision is intentionally narrow and reflects the English courts' strong policy in favour of the finality of arbitral awards.
Accordingly, the mere use of AI would not, without more, amount to a serious irregularity. Nor would an isolated legal error or inaccurate citation ordinarily justify intervention.
However, the analysis may differ where fictitious authorities form a material part of the tribunal's reasoning. If reliance on AI-generated hallucinations demonstrates that the arbitrator failed to verify legal authorities or effectively delegated aspects of the decision-making process, an applicant could argue that the tribunal failed to conduct the proceedings in accordance with its adjudicative obligations.
Whether such a challenge would succeed will depend on the particular facts. Nevertheless, the Québec decision illustrates the type of circumstances in which the distinction between permissible AI assistance and impermissible delegation may become legally significant.
AI in arbitration: emerging safeguards
Recognising both the opportunities and risks associated with AI, arbitral institutions have begun issuing guidance on its responsible use.
The CIArb Guidelines encourage parties to address AI use at the outset of proceedings and recommend that tribunals record AI-related decisions in procedural orders. They also caution against using AI for legal analysis, legal research, interpretation of evidence or the application of law to facts without independent verification.
These safeguards are designed to preserve the arbitrator's personal responsibility for the award while allowing AI to improve efficiency in appropriate administrative and drafting tasks.
A model for responsible AI use
Rather than prohibiting AI altogether, some institutions are exploring models that combine AI assistance with meaningful human oversight.
In 2025, the International Centre for Dispute Resolution (ICDR) announced that AI-assisted arbitrators would be available for certain document-only construction disputes. Under this model, AI prepares a draft award using structured prompts, but a human arbitrator reviews the reasoning, verifies the authorities, makes any necessary revisions and assumes full responsibility for the final award .
The model reflects an emerging consensus that AI can enhance arbitration, provided accountability remains firmly with the human decision-maker.
Conclusion
The Québec decision should not be understood as a rejection of AI in arbitration. Instead, it reinforces a long-established principle: the responsibility for legal reasoning, factual analysis and the final decision rests with the arbitrator alone.
As AI becomes increasingly embedded in legal practice, courts are likely to distinguish between AI used as an efficiency tool and AI used as a substitute for independent judgment. The Québec decision suggests that where that line is crossed, the consequences may extend beyond reputational damage to the validity and enforceability of the award itself.
