When AI Enters the Arbitrator’s Room: A Warning About Judgment, Delegation, and Trust

Artificial intelligence is rapidly becoming part of the legal profession. Lawyers use it to summarize documents, organize evidence, test arguments, clean up drafts, and identify issues. Courts are beginning to address its use in pleadings and briefs. Arbitrators and mediators are also confronting the same practical question: where can AI help, and where does it threaten the integrity of the process?

A recent decision from the Quebec Superior Court may be one of the first reported decisions in the world to set aside an arbitration award because of the arbitrator’s apparent use of generative AI.

The decision should get the attention of every arbitrator, mediator, advocate, and client who participates in private dispute resolution.

The issue was not simply that AI may have been used. The court did not say that every use of artificial intelligence by an arbitrator is improper. The problem was deeper. The award relied on legal authorities that did not exist. Those authorities were not peripheral. They were central to the arbitrator’s reasoning. The court concluded that the arbitrator had effectively delegated part of his decision-making function and abdicated his responsibility to verify the result.

That is the real lesson.

AI can assist a decision-maker. It cannot become the decision-maker.

The Case in Broad Terms

The underlying dispute involved a claim for retroactive compensation arising out of agreements governing intermediate housing resources in Quebec’s health and social services system. The arbitrator dismissed the claim on timeliness grounds, finding that the applicable disagreement procedure had not been followed within the required time.

The claimants then sought to annul the award.

They made two principal arguments. First, they argued that the award violated public order because it enforced a contractual time limit shorter than the statutory limitation period. Second, they argued that the arbitration procedure had not been respected because the award relied on non-existent doctrine and case law, suggesting that the award had been drafted with the assistance of artificial intelligence.

The court rejected the first argument. That part of the decision is important because it reinforces the limited role courts generally play when reviewing arbitral awards. An annulment application is not an appeal. A court is not supposed to reweigh the merits, correct ordinary legal errors, or decide whether the arbitrator’s reasoning was persuasive. Arbitration depends on finality, and courts are rightly cautious about converting award review into ordinary appellate review.

But the court accepted the second argument.

The award cited a scholarly article that could not be found. It cited appellate and trial-level decisions that did not exist or whose numerical citations led to unrelated cases. It cited an arbitral award that the relevant legal database confirmed did not exist. These were not minor citation mistakes. They were the only doctrinal and case-law authorities used to support the arbitrator’s legal conclusions.

That changed the analysis.

The court concluded that the issue was not whether the arbitrator had made a correct or incorrect legal ruling. The issue was whether the agreed arbitral procedure had been respected. Because the parties had chosen a specific arbitrator to decide their dispute, they were entitled to a decision made by that arbitrator — not by an opaque technological tool generating false legal support.

Why the Arbitrator’s Role Matters

Arbitration is built on consent. Parties agree to remove their dispute from court and submit it to a private decision-maker. That decision-maker may be selected because of subject-matter knowledge, legal judgment, experience, efficiency, temperament, availability, or trust.

That choice matters.

When parties choose an arbitrator, they are not simply choosing a signature block at the end of an award. They are choosing the person who will hear the evidence, assess the arguments, deliberate, apply the law, and explain the result. The arbitrator’s judgment is the product being purchased. It is also the foundation of the award’s legitimacy.

The Quebec decision emphasizes that point. The court focused on the parties’ autonomy in choosing their decision-maker, the importance of reasoned awards, the secrecy of deliberations, and the rule against delegating adjudicative authority.

Those principles are not technicalities. They go to the core of arbitration.

A reasoned award is not just a written explanation attached after the real decision has already been made. The act of writing reasons can test the decision itself. It forces the arbitrator to identify the issues, confront weaknesses, evaluate competing arguments, and determine whether the initial conclusion still holds. Anyone who has written decisions, briefs, bench memoranda, or serious legal analysis understands this. The reasoning process can change the result.

That is why delegating the reasoning process can become delegating the decision.

AI Assistance Versus AI Delegation

The decision should not be read as a blanket rejection of artificial intelligence in arbitration. The court expressly recognized that technological tools can improve access to justice and that legal professionals increasingly use large language models for tasks such as summarizing documents, identifying trends, transcribing recordings, refining text, and assisting with research.

The line is not between technology and no technology.

The line is between assistance and delegation.

AI may be used as a tool, but the arbitrator must remain responsible for the decision. That means the arbitrator must independently assess the facts, law, arguments, and result. It also means that any authorities used in the award must be verified. If AI generates a proposed case citation, article, quote, proposition, or analytical structure, the arbitrator cannot simply accept it because it looks plausible.

That is especially important because generative AI is built to produce fluent text, not guaranteed truth. Hallucinated legal authorities can look real. They may have realistic party names, court names, docket numbers, dates, and summaries. They can be convincing precisely because they mimic the language of legal writing.

That is what makes them dangerous.

A fake citation in a lawyer’s brief is serious. A fake citation in an arbitral award is more serious still. The award is the decision. If the award rests on law that does not exist, the parties may reasonably question whether the arbitrator actually performed the adjudicative function the parties selected him to perform.

The Court’s Careful Limitation

One of the most important aspects of the Quebec decision is that the court did not announce a rule that every erroneous citation or every use of AI requires annulment.

That limitation matters.

Legal writing sometimes contains mistakes. A citation may contain a typo. A quotation may need correction. A case may be cited for a proposition that is too broad. An arbitrator may use dictation software, translation assistance, research support, clerical help, or drafting tools without compromising the award.

The court recognized that there may be cases where AI use is minimal, where an error is peripheral, or where the problem does not affect the integrity of the process or the outcome.

This case was different.

The non-existent authorities were central to the reasoning. They were the legal support for the award’s conclusion. The court found that the problem was significant, likely affected confidence in the result and the arbitration system, and probably had a meaningful impact on the outcome. That justified annulment.

That careful limitation is important for arbitration users. The decision is not anti-technology. It is pro-accountability.

Potential Ramifications for Arbitration

The decision may have significant consequences beyond Quebec.

First, it gives parties a new way to think about challenges to arbitral awards. Courts are reluctant to disturb arbitration awards, and they should be. But if an award appears to rely on hallucinated authorities, the issue may not be framed as ordinary legal error. It may be framed as a failure to follow the agreed arbitral procedure, an improper delegation of decision-making authority, or a defect affecting the integrity of the process.

Second, the decision may increase post-award scrutiny. Counsel receiving an unfavorable award may now be more likely to verify the authorities cited in the decision, particularly where the reasoning seems generic, unusually abstract, or supported by unfamiliar citations. That does not mean every award should be attacked. It does mean that citation verification may become part of post-award due diligence.

Third, arbitrators may need to adopt clearer AI practices. At a minimum, arbitrators who use AI tools should have procedures to ensure that no confidential information is improperly entered into public systems, no generated authority is used without verification, and no reasoning is accepted without independent judgment. Arbitrators may also need to consider whether, when, and how to disclose AI use, particularly if the use goes beyond formatting, proofreading, or administrative assistance.

Fourth, arbitral institutions may respond with rules, protocols, or guidance. Some may require disclosure of AI use. Others may prohibit certain uses. Some may permit AI for administrative or drafting support but forbid its use for independent legal reasoning, credibility assessment, damages analysis, or merits decision-making. The challenge will be to regulate AI without freezing useful technology in place.

Fifth, the decision may affect arbitrator selection. Parties may begin asking prospective arbitrators about their use of AI, confidentiality safeguards, citation-verification practices, and drafting process. That should not be seen as offensive. It is a practical question about how the decision will be made.

Practical Lessons for Arbitrators

For arbitrators, the lesson is straightforward: AI output must never be treated as authority.

If AI is used at all, it should be used with discipline. It may help organize information, refine language, or identify issues for further review. But every legal proposition must be checked against reliable sources. Every citation must be verified. Every quotation must be confirmed. Every conclusion must be the arbitrator’s own.

Confidentiality is equally important. Arbitration often involves sensitive business records, personnel information, trade secrets, medical information, financial data, or settlement communications. Uploading that material into an AI tool may create confidentiality risks. Arbitrators should understand the tool being used, where data goes, whether it is retained, and whether it may be used to train future systems.

Most importantly, arbitrators must preserve the human act of judgment. The parties selected an arbitrator, not a language model.

Practical Lessons for Counsel and Clients

For counsel, this decision suggests several practical steps.

Before arbitration begins, consider addressing AI use in the arbitration agreement, procedural order, or preliminary conference. The parties may want to define what uses are permitted, what uses are prohibited, whether disclosure is required, and how confidentiality will be protected.

During the arbitration, counsel should verify their own AI-assisted work before submitting it. Courts and tribunals have little patience for fake authorities, whether they appear in a brief, exhibit summary, or legal memorandum.

After an award is issued, counsel should review the authorities that matter to the result. That does not mean looking for harmless typographical errors. It means confirming that the authorities central to the award actually exist and support the propositions for which they are cited.

For clients, the practical point is even simpler. Arbitration depends heavily on trust. Parties agree to accept a private decision because they trust the process, the rules, and the decision-maker. When AI is used carelessly, that trust can be damaged quickly.

A Warning, Not a Rejection

The Quebec decision is likely to become an important reference point in the ongoing discussion about artificial intelligence in arbitration. It may be the first decision of its kind, but it almost certainly will not be the last to confront these issues.

The better reading of the decision is not that AI has no place in arbitration. The better reading is that AI must remain in its place.

Used carefully, AI may help lawyers and neutrals work more efficiently. Used carelessly, it can generate false authority, compromise confidentiality, distort reasoning, and undermine confidence in the award. Used as a substitute for judgment, it threatens the foundation of the arbitral process itself.

The promise of arbitration is not merely that disputes will be resolved privately or efficiently. The promise is that the parties will receive a decision from the neutral they selected, based on the record, the law, and that neutral’s independent judgment.

That promise cannot be delegated.

Nationwide ADR® provides arbitration, mediation, and other dispute-resolution services designed to bring structure, judgment, and clarity to demanding disputes. Unlocking Solutions for Demanding Cases®.