AI in arbitration: a tool for assistance, not decision-making

Generative AI has become part of everyday legal practice ‒ and has found its way into arbitration. It can be used in many ways: to summarise information, prepare documents and even assist with drafting a decision. As is often the case with new technologies that remain lightly regulated, however, questions quickly arise as to where the limits lie. The Superior Court of Québec (Cour supérieure) has now provided an answer in its decision of 22 April 2026: AI may assist, but it must not decide (2026 QCCS 1360). Given the limited legal analysis of the issue to date, the decision also offers useful guidance for parties to arbitrations seated in Germany and to recognition or enforcement proceedings.

The case: AI at the heart of the decision to set aside the award

As far as is known, the Canadian decision is the first in which a court has set aside an arbitral award because of the use of AI. In the underlying arbitration, the sole arbitrator apparently used generative AI to draft his arbitral award. This was evident from the grounds given: key citations from case law and academic commentary on which the award relied either could not be traced or in fact referred to entirely different cases. The sources were clearly hallucinated – a well-known risk associated with generative AI. However, the arbitrator appears to have done more than introduce AI-generated citations. The court found that he had relied not on his own independent judgement but on AI when reaching the decision as a whole.

The unsuccessful party successfully challenged the award on, among other, the following grounds:

• The arbitrator had in effect delegated his decision-making authority to an AI system.
• By relying on non-existent authorities, he had deprived the parties of a fair and transparent process.

Clear red line: arbitrators must not delegate decision-making to AI

In its decision, the court emphasised three fundamental principles that also apply under German law and the New York Convention, which governs international arbitral awards:

• Personal responsibility for deciding the dispute and the right to appoint an arbitrator

As a rule, the parties appoint an arbitrator themselves (for arbitrations seated in Germany, this is reflected in section 1035 of the German Code of Civil Procedure (Zivilprozessordnung). The arbitrator must decide the dispute independently and personally. They cannot delegate that decision-making authority to another person or to an AI system.

• Confidentiality of the arbitral tribunal’s deliberations

An arbitrator must preserve the confidentiality of deliberations. Once the matter is ready for decision, the arbitrator may not discuss the substantive assessment of the case with third parties. This confidentiality is intended to ensure that the arbitral award reflects the arbitrator’s own convictions not those of an external third party or external system. According to superior court decisions, this principle also applies in German arbitration law, although it is not expressly codified, as it is for state judges under section 43 of the German Judiciary Act (Deutsches Richtergesetz).

• Arbitral tribunal must give its own independent reasons

The reasoning behind a decision is at the heart of adjudicative work and ensures confidence in the decision. The parties are entitled to expect that the arbitrator they have appointed will conduct their dispute, decide it and give the reasons for the decision personally. Unless the parties have agreed otherwise, arbitral proceedings seated in Germany require those reasons to be given in writing (section 1054(2), first sentence of the German Code of Civil Procedure). Hallucinated citations undermine that reasoning function and, with it, the integrity of the proceedings.

Applying these fundamental principles, the Canadian court reached the following finding, which led to the setting aside of the arbitral award: “The balance of probabilities therefore leads to the conclusion that the arbitrator’s authority was delegated and that he abdicated his role in reviewing the outcome.” (Judgment of the Superior Court of Québec dated 22 April 2026, para. 114)

An application to set aside under section 1059 of the Code of Civil Procedure in respect of arbitral proceedings seated in Germany, or recognition and enforcement proceedings under section 1061(1) of the Code, would likely lead to a comparable outcome. As explained above, delegating the decision-making process to AI and relying, in central parts of the award, on hallucinated citations would breach comparable fundamental procedural principles of German arbitration law. 

The decision thus opens up a new dimension in the review of arbitral awards. The Canadian court further held that no expert evidence was required to establish the use of AI. The applicable standard was the civil standard of proof – the balance of probabilities – and unverified AI-generated material, such as fabricated legal sources or inconsistent reasoning, was in itself sufficient to support the inference that AI had been used.

AI in arbitration: what remains permissible and what needs to be borne in mind?

The decision is not a ban on AI. Rather, the Canadian court recognised that technological tools, including AI, may assist lawyers and arbitrators in their work. In particular, the following remain permissible:

• Using AI for research, structuring arguments, summarising documents, suggesting wording or providing translations.
• Using other support resources, such as tribunal secretaries or AI systems, provided that the arbitrator:
  • makes the decision personally;
  • develops the reasons personally;
  • verifies sources independently; and
  • preserves the confidentiality of deliberations and of the proceedings.

The dividing line is therefore clear: AI as a tool: yes. AI as a decision-maker or co-decision-maker: no. Many international arbitral institutions and organisations take the same approach and have already issued guidance on the use of AI in arbitral proceedings, including:

• the “Guidelines on the Use of Artificial Intelligence in Arbitration” published by the Silicon Valley Arbitration & Mediation Centre in 2024;
• the “Guide to the use of artificial intelligence in cases administered under the SCC Rules” published by the Stockholm Chamber of Commerce (SCC) Arbitration Institute in 2024;
• the “Guidance on Arbitrators’ Use of AI Tools” published by the American Arbitration Association – International Centre for Dispute Resolution in 2025;
• the Chartered Institute of Arbitrators’ “Guideline on the Use of AI in Arbitration” published in 2025; and
• the “Note on the Use of Artificial Intelligence in Arbitration Proceedings” published by the Vienna International Arbitration Center in 2025.

In addition, there are further risks associated with the use of AI, which the Canadian court highlighted. Those risks also apply to non-Canadian cases:

• Hallucinations: AI may invent case law and academic commentary, presenting them as supposedly genuine sources.
• Inadequate judgement in discretionary matters: AI’s rule-based logic may be incompatible with discretionary assessments that require consideration of social values, the parties’ individual circumstances or other contextual factors.
• Bias and black box: The inner workings and training data of many models are not transparent. This makes it more difficult to assess their outputs for possible bias.
• Confidentiality: Uploading sensitive arbitral materials to public AI systems may breach duties of confidentiality and data privacy obligations.
• Confidence in the arbitral process: If parties believe that a machine, rather than the tribunal, decided the dispute, confidence in and acceptance of the process and the award may be undermined.

Conclusion: allow AI consciously, prevent delegation

The Canadian court’s decision is a leading authority on the judicial treatment of generative AI in arbitration. It accepts AI as a form of technical assistance but rules out delegating the core adjudicative function – independent decision-making and the provision of reasons – to a machine. The same applies to German parties: AI has a place in modern arbitral practice, but only within clearly defined guardrails.

Those guardrails can be established through careful procedural planning. At the outset of the arbitration, parties should clearly communicate their expectations regarding the use of AI to the tribunal. Conversely, tribunals would be well advised to raise any intended use of AI assistance proactively and agree the appropriate parameters with the parties, much as they would when appointing or using a human tribunal secretary.