An Algorithm Is Not An Arbitrator: The Consequences of Delegating Decision-Making To Artificial Intelligence - A Quebec Case Study

Few workplaces have yet to feel the presence of Artificial Intelligence (AI) in at least some form, and the justice system is no different. But, while some AI tools have been deployed effectively to reduce the time workers spend on tedious tasks, there can be serious consequences if the line between delegating tasks and delegating responsibility is crossed.

Recently, the Quebec Superior Court took the rare step of setting aside a labour arbitrator’s decision because his reasoning consisted entirely of doctrinal and case law references that were the product of AI hallucinations. For those less familiar with the term “AI hallucination”, it essentially means that an artificial intelligence platform has produced a response or information, such as a case law citation or passage, that sounds confident or plausible but is actually false, contextually misleading, or entirely made up. 

The Superior Court’s Decision – while not dismissing the potential value of AI tools outright – is another example in an increasingly long line of instances where indiscriminate AI use has resulted in profound legal problems.

The Case Before the Arbitrator.

In an article for Justice-Quebec.ca, Maxime Gagné describes the matter at the centre of the arbitration as “a fairly ordinary dispute between a healthcare provider and Quebec’s public health network”. The Centre de Santé Osman, an intermediate residential care resource, made a claim totalling $1,225,000 against a former health agency now incorporated into Santé Québec. The amount claimed was for residential placements declared available during a three-year period (but for which no users had been referred). 

Although the health agency communicated a clear and final refusal to pay in November 2021, the Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) waited almost three years to transmit a notice of its disagreement for arbitration. This delay exceeded a strict 90-day period to file disputes in a binding agreement between the parties.

The ARIHQ and Centre de Santé Osman contended the three-month deadline was contrary to article 2884 of the Civil Code of Québec which prohibits modifying limitation periods by contract. Québec’s general limitation period is three years.

Arbitrator Michel A. Jeanniot issued his decision on August 8, 2025, confirming that the contractual limitation was valid.

Dead End Citations.

When the lawyers for the ARIHQ and Centre de Santé Osman reviewed sources Jeanniot had cited to support his decision, they encountered a dead end. And then another. And another.

Three of the neutral citations for case law included in the decision referred to three completely different judgments.

To make matters worse, as Quebec Superior Court Justice Martin F. Sheehan noted, the three citations were critical to Jeanniot’s reasoning. These citations, alone, were the sole authorities employed to provide the legal doctrine and jurisprudential basis for the arbitrator’s award. All other citations appeared in summaries of the parties’ positions. 

No Possibility for Appeal, But an Application to Annul.

When parties agree to binding arbitration, it generally precludes the right to appeal the arbitrator’s decision. However, under certain circumstances, an applicant can file an application for the decision to be annulled. 

In Quebec’s Code of Civil Procedure, article 646 narrowly prescribes five circumstances which applicants can invoke when applying to a court to set aside an award:

• incapacity or invalidity of the arbitration agreement
• failure to follow the applicable method and procedure for appointing the arbitrator and for the arbitration proceedings
• inability of a party to assert its case
• award dealing with a dispute not contemplated by the agreement 
• award contrary to public order

Upon review of the facts presented to support the application, Justice Sheehan dismissed the applicant’s first argument that the award is contrary to public order because it upheld a contractual deadline shorter than the legal limitation period. He explained that even an error in law involving a public order provision is not sufficient grounds for annulment. Rather, the result of the award, not the reasoning, must offend fundamental principles of public order for this condition to apply.

Delegatus Non Potest Delegare – One Who Receives Power Cannot Delegate It.

A second argument – that the arbitration procedure had not been respected due to the arbitrator allegedly delegating his decision-making power to an AI tool – was accepted. 

The principle that power bestowed upon an arbitrator cannot be delegated is a cornerstone principle of both administrative and arbitration law. Justice Sheehan noted that in this particular case parties had to mutually agree to an arbitrator from a list of 10 candidates. Presumably parties had their own reasons for making their choice – from an arbitrator’s availability to hear the dispute to personal qualities such as experience, expertise, or reputation. Justice Sheehan stated that these parties were entitled to expect their chosen arbitrator would render a decision as opposed to a third party or “an algorithm”.

Justice Sheehan also signalled that when it comes to written reasons for a decision, the drafting process is an integral component which requires reflecting on source material. Finally, he pointed to article 644 of the Code of Civil Procedure which requires an arbitrator to preserve confidentiality of deliberations. He suggested that sharing details with an external AI tool is a potential breach of this duty.  

A central unifying theme of these points is the arbitrator’s responsibility to draft his own written decisions. While an arbitrator may employ certain tools or consult specific people or sources when drafting a decision, this must be done in a way that protects the arbitrator’s own responsibilities, duties and, particularly, the integrity of the process.

The Risks of AI.

Justice Shaheen’s decision called attention to five key risks of using AI in judicial settings:

1. Hallucinations – AI, and specifically large language models (LLMs), are designed to search accumulated patterns in order to predict what a user is hoping to see or create based on their prompts. As Richard Murphy explains in an Enago article, “They generate plausible text given a prompt, but they do not ‘retrieve’ verified bibliographic records in the way a database does. When asked for citations, models may invent titles, Digital Object Identifiers (DOIs), or journal names that fit learned patterns. Retrieval-augmented approaches (RAG) can reduce this risk but do not eliminate it.” 
2. AI rigidity versus human discretion. AI is incapable of making subjective assessments based on immersive lived experience within a community. Structured patterns cannot adequately account for how particular context will shape the way community values are understood and applied. 
3. Bias. Both AI and humans can demonstrate bias. However, while humans can be questioned on how they came to certain conclusions, many AI developers decline to share source code which could be used to detect biases reproduced from a model’s training data. As a result, AI has a “black box” problem.
4. Lack of confidentiality. Information submitted into an AI tool is not considered private or privileged, as I noted in a recent blog post. Once this information becomes integrated into an AI database, it could be disclosed and breach the duty for confidentiality of deliberation.
5. Public confidence. The public mood toward AI has soured considerably in recent months. If parties to an action discover that a machine has rendered a decision in place of a human - or even believe this could be a possibility – trust in the justice system would be shaken. 

Not a Blanket Ban, But a Plea for Caution.

To err is human. With or without the aid of AI, people have been known to unwittingly introduce errors in their submissions to court (or in their judgments). Establishing an absolute rule to set aside any decision containing evidence of AI would mirror one of the critiques Justice Shaheen had of the technology itself: rigidity when context calls for flexibility. 

Justice Shaheen said a proportionality test should apply when considering annulment in such situations. This test would evaluate:

• the nature of the breach
• the severity of its impact on the procedure’s integrity
• the actual impact of the breach on the award itself

The case before him passed the threshold for annulment because an AI fabrication was the foundation of and determinative factor for the award, and this breach was likely to be detrimental to the parties’ confidence in the entire arbitration regime, not simply this particular award.

As Gagné notes in his article, while the case before Shaheen shares some similarities with other instances of AI fabrications in court proceedings, “the originality of the Sheehan decision lies in the fact that it targets a decision-maker. And it lays down a principle that extends well beyond the case at bar. For arbitrators, administrative judges, and other adjudicators, the message is clear: the tool may assist, but it can never replace the deliberative function”.

Conclusion.

It’s been said that a tool is only as good as the person who wields it. If used indiscriminately, a chainsaw can cut off a limb. When handled attentively and with the utmost safety in mind, it can save time and energy compared to swinging an axe to fell a tree.

As a long-time proponent of employing emergent technologies to enhance the work product of the legal profession, it pains me to see how misuse of these tools can engender greater suspicion of AI and of people who deploy it thoughtfully. As these cases mount, some people may even outright dismiss the technology entirely.

This case, and others like it, should serve as a reminder to everyone in our profession to take time to learn how to use new technology properly. We must carefully consider how to use technology to support us as we fulfill our professional responsibilities, rather than a means to abdicate them.