An arbitrator citing AI-generated sources that never existed – and a court setting aside the arbitral award as a result: what sounds like a textbook example from a legal tech seminar has become reality in Canada. So far, sanctions for AI hallucinations have mainly targeted lawyers and experts. Now, an arbitral tribunal is in the spotlight for its use of AI. The decision could send a signal well beyond Canada – including to Germany.
ARIHQ v. Santé Québec – 2026 QCCS 1360
On 22 April 2026, the Québec Superior Court (Cour supérieure) handed down a decision setting aside an arbitral award on the grounds of improper use of generative AI.
What happened?
The arbitration concerned a claim by a clinic for approximately CAD 1.2 million against a public healthcare institution in the province of Québec for services provided between 2019 and 2022. The arbitral tribunal dismissed the claim for failure to comply with a preclusive time limit. In November 2025, the claimants applied to annul the award, arguing that there were strong indications the reasoning had been drafted using a hallucinating AI: all legal sources on which the arbitral award was based were the result of a hallucinating AI.
Grounds for setting aside the award
The Canadian court confirmed the claimants’ allegations and annulled the award pursuant to Article 646 no. 3 of the Québec Code de procédure civile. This provision – similar to Article V (1) (d) of the New York Convention – allows for annulment where the applicable arbitration procedure was not observed. The court focused on the intuitu personae nature of arbitration: an arbitrator is chosen precisely for their personal expertise and judgment. Delegating the substantive reasoning to an AI tool without sufficient review, violates the principle of delegatus non potest delegare and breaches the confidentiality of deliberations, the court found. The court based its decision on the following three pillars:
Principle of party autonomy – selection of the arbitrator: Parties entering into an arbitration agreement are, in the court’s view, entitled to have the selected arbitrator render the decision.
Reasons for the decision: Citing the case law of the Supreme Court of Canada, the court observed that providing reasons promotes better decisions and enables parties to understand the outcome and evaluate their options for appeal.
Confidentiality of deliberations: Arbitrators must, in the court’s reasoning, refrain from communicating with third parties about the case once accepted for arbitration. This ensures that the award reflects solely the arbitrators’ views. Entering confidential information into an AI tool therefore breaches this duty.
While the court emphasised that not every use of AI would lead to annulment, it identified five risks associated with AI in decision-making:
Hallucinations: Large Language Models can invent sources that appear deceptively genuine.
Lack of human judgment: AI may fail to adequately account for ‘human' values and specific circumstances.
Bias: Training data may skew AI outputs.
Confidentiality: Uncertainty remains about how input data is processed and stored.
Lack of trust: Decisions by ‘disembodied’ machines could undermine public confidence in the justice system.
The court concluded that it is therefore crucial that responsibility for the reasoning remains with the decision-maker. Whether the use of AI could lead to annulment depends on the circumstances of each case, the court stated. In this case, the breach was significant: the award relied on an academic article, three court decisions, and an arbitral award – none of which actually existed. These AI-generated fabrications weren’t just footnotes; they formed an essential part of the arbitrator’s reasoning and were the sole legal authority cited for the entire award.
Implications for German arbitration practice
How would a German Higher Regional Court (Oberlandesgericht) assess a comparable case where the place of arbitration is in Germany? Under Section 1059 (2) no. 1 (d) of the German Code of Civil Procedure (ZPO), an arbitral award may be set aside if “the formation of the arbitral tribunal or the arbitral proceedings was not in accordance with a provision of the present Book or with an admissible agreement in place between the parties, and that this presumably has had an effect on the arbitral award”. This provision, like Article 646 no. 3 of the Québec Code de procédure civile, is based on Article 34 (2) (a) (iv) of the UNCITRAL Model Law (similar to Article V (1) (d) of the New York Convention). For this reason alone, the Canadian decision is likely to send a strong signal for German arbitration proceedings.
The use of AI is often compared to that of a human assistant. Even in the analogue world, drawing the line between permissible and impermissible use of assistants is complex and depends heavily on the parties’ agreement. This applies all the more in the digital world. From a German perspective, too, the decisive factor is likely to be that the final decision-making remains the sole prerogative of the arbitrator. If arbitrators merely adopt content without forming their own independent assessment of the facts and legal position, they are relinquishing their decision-making authority. An award that is essentially based on unverified AI content embodies precisely this lack of independent engagement with the facts and legal issues of the case.
This principle is also emphasised in the guidelines of several arbitration institutions:
Silicon Valley Arbitration & Mediation Centre, Guidelines on the Use of Artificial Intelligence in Arbitration (2024): “The use of AI tools by arbitrators shall not replace their independent analysis of the facts, the law and the evidence.”
SCC Arbitration Institute, Guide to the use of artificial intelligence in cases administered under the SCC Rules (2024): “AI tools may be used to support arbitral decision-making but cannot replace it. Arbitral tribunals cannot delegate the decision or the reasoning leading to it to anyone or anything.”
American Arbitration Association, Guidance on Arbitrators’ Use of AI Tools (2025): “When using AI tools for legal research or evidence analysis, arbitrators should ensure that their decisions reflect their independent evaluation and reasoning.”
Chartered Institute of Arbitrators, Guideline on the Use of AI in Arbitration (2025): “[T]he Tribunal should avoid delegating any tasks to AI tools, such as legal analysis, research and interpretation of facts and law, or the application of the law to the facts, if such use could influence procedural or substantive decisions.”
Future cases are unlikely to be as straightforward. Much like working with a human assistant, there is a broad spectrum between AI-supported assistance and de facto delegation of decision-making. The line becomes especially hard to draw when AI is used to prepare the substance of a case. Fact summaries, legal research, or suggested wording for contentious issues all risk unconsciously steering decisions in a certain direction (‘automation bias’).
Conclusion
The Québec decision demonstrates that the intense debate surrounding AI in arbitration has reached the practical stage: not only lawyers and experts, but also arbitrators are now being held accountable for the consequences of uncontrolled AI use. In this case, the numerous hallucinations left no doubt as to the impermissible delegation. For counsel, the most challenging cases going forward will be those where the boundary between legitimate AI assistance and improper delegation is unclear – or where AI misuse is harder to detect because there are no obvious hallucinations in the arbitral award.
Forward
