What happens when a judge, arbitrator, or other legal decision-maker issues a decision that appears to be AI-generated?

ARIHQ v CIUSSS, 2026 QCCS 1360.

Decided April 22, 2026

This case, a decision on a labour arbitration in Quebec, may not appear relevant to criminal law at first glance. But the decision deals with an increasingly relevant issue for all areas of law, regardless of location or severity of the matter.

What happens when a judge, arbitrator, or other legal decision-maker issues a decision that appears to be AI-generated?

To provide some background (without getting too deep into Quebecois labour and healthcare arbitration), this case involved a dispute between ARIHQ, which represents community-based care facilities for seniors and adults suffering from a loss of autonomy, and the CIUSSS, which manages hospitals and healthcare facilities in southern Montreal. The parties had an established dispute resolution agreement, which stated the parties must initially try to resolve disputes amicably. If those efforts failed, a notice of disagreement had to be filed within a 90-day deadline for litigation to proceed.

A dispute between the parties first arose in 2021 regarding non-payment from the CIUSSS. Meetings held in September of 2021 to attempt an amicable resolution were unsuccessful. However, ARIHQ did not meet the 90-day deadline for filing a notice of disagreement, filling on August 1, 2024. CIUSSS responded on August 28, saying that the notice was filed past the deadline and objecting to any further action on this matter. Regardless, ARIHQ filed a request for arbitration on September 13, 2024. The arbitrator upheld CIUSSS’s objections in his decision on August 8, 2025. ARIHQ appealed that decision, bringing us to the case at hand.

At the core of this appeal was ARIHQ’s claim the arbitrator had failed to comply with the dispute resolution agreement, specifically by deferring their decision-making authority to Artificial Intelligence. Though the respondents raised that there was no rule in the Quebec Code of Civil Procedure that specifically barred the use of AI in crafting a decision, the Court looked deeper to consider the potential issues that AI-generated decision-making could cause.

In finding that the decision was AI-generated (a necessary first step) the Court considered the references to non-existent and “delusional” cases that were included in the decision. The Arbitrator had cited six cases in the core of the decision, five of which were non-existent. The one real case cited by the Arbitrator was misrepresented and did not actually support the Arbitrator’s decision. The Court found this sufficient to meet the civil standard of balance of the probabilities, as given the non-existent caselaw references, the decision was more likely than not to have been AI generated. This case still leaves the question open as to how this would apply to the “beyond a reasonable doubt” standard required in the criminal law.

The QCSC took this opportunity to discuss how the use of AI to write legal decisions may violate legal principles. First, the court acknowledged that the agreement between the parties required them to mutually agree on the arbitrator used from a list of approved arbitrators. It was a violation of this agreement to outsource the decision to a computer program. It also undermined the importance of a decision-maker writing reasons for their decisions. The QCSC offers a quote from the Supreme Court in Baker v Canada (Minister of

Citizenship and Immigration), [1999] 2 SCR 817 to illustrate this point:

“[A]nyone entrusted with a judicial function knows full well that it is by attempting to provide reasons for a decision based on a first impression that one often discovers that the decision is not the right one.”

Allowing an AI program to write a decision, potentially based on that incorrect first impression, would undermine any expectation of fairness in our legal system. The Canadian Judicial Council has explicitly forbidden judges from delegating their roles as decisionmakers, and while that rule was presumably intended to apply to human delegates, the Court considered this relevant in considering AI delegation as well.

Privacy and confidentiality concerns were also highlighted, as the use of an AI program in legal decision-making inherently requires uploading sensitive information. The QCSC noted that the companies responsible for these AI programs are often unwilling to discuss their source code, and therefore users cannot assume that information they upload is being kept secure. The Court also considered the “opaque nature” of AI program as a factor that allows biased reasoning to develop unchecked. These biases may or may not be intentional and may come from either the programmers or the data sets the AI programs are trained on. There is virtually no way for a user to “look under the hood” and assess what’s really going on. The QCSC considered this inability to properly assess and correct bias in AI programs as a key weakness when compared to human decision making.

Continuing that theme, the QCSC also considered the potential lack of human discretion and values in an AI-generated decision, coming to the conclusion that the rigidity of a computer program cannot be expected to match the complex and flexible nature of the human brain, and the public’s faith in the legal system would be justifiably undermined if AI programs were allowed to usurp the role of judge. The Court summarized this point at paragraph 90: “Justice requires that decisions be perceived as fair, humane, and reasoned. If the public believes that a disembodied or opaque machine made a decision, rather than a human being, confidence in the judicial system could be undermined.” Despite this, the Court does not expressly rule that all decisions made using AI will be overturned the same way. The QCSC leaves open the possibility of decision-makers using AI in “less significant” ways, suggesting that future appeals may turn on the impact of the AI use on the “integrity of the proceedings” and its effect on the decision being made.  However, this decision has offered a thorough set of arguments against the use of AI in this context, which will only become more relevant moving forward.