In early September 2026, the chief justices of Quebec's Court of Appeal, Superior Court, Court of Québec and municipal courts issued common guidelines on judges' use of generative AI. Their central line is that judging is a human act that cannot be delegated. The guidelines bar AI from legal reasoning, from analysing facts, from assessing evidence and credibility, from determining outcomes, and from drafting those parts of a decision, according to Global News. The principle is right. The harder question is whether a document made mostly of prohibitions equips judges to work safely in a world where AI is already in their inboxes.
The strongest case for the Quebec approach
The case for strictness is substantial. Global News reports that the guidelines followed a March 2026 La Presse report that a Quebec court decision cited case law that does not exist, raising the possibility that generative AI was involved. A court's authority rests on the public's confidence that a named human weighed the evidence and answers for the result. A fabricated citation in a judgment is not a typo. It damages that confidence in a way a flawed brief from a lawyer does not.
The courts also give a practical reason. As summarised by Canadian Lawyer, the guidelines say no authorised, secure AI tool has been deployed in the four courts, and commercial tools are largely opaque. Judges handle confidential material, including draft judgments and case notes. The guidelines tell judges not to feed that material to AI tools, and, until an approved institutional tool exists, to make sure their settings keep inputs from training the system. Where a judge is unsure, the guidance is to refrain from using generative AI.
There is nothing here a free-speech or pro-innovation reader should resent. Courts are not markets. Nobody has a right to an algorithmic judge, and parties have a strong interest in a decision-maker who can be cross-examined by the record, appealed, and held to account.
What the guidelines allow
The text is not a blanket ban. Global News reports that judges may use AI, under human control, to revise, rephrase, correct or translate text they have already written, for certain administrative tasks, and to locate information within specified sources. All output must be independently verified. Judges stay fully responsible for the reasoning and for the accuracy of every fact, quotation and authority in their decisions. Canadian Lawyer's account also describes an allowance for purely descriptive summaries.
That is a defensible floor. It leaves the thinking with the judge and permits assistance that resembles a better spellchecker.
Where the critique lands
In a 15 September 2026 commentary in Canadian Lawyer, Toronto lawyer Dmitry Shniger says: "I agree with the substance, but I'm very disappointed with the form." His complaint is that the document lists prohibitions and risks, which leaves a judge two options. Either abstain whenever the position is unclear, which the guidelines expressly recommend, or conclude that compliant use is impossible.
Three of his points deserve weight.
- The summary allowance may be hollow. Shniger argues that AI's real strength is spotting inconsistencies across a record, which a purely non-judgmental summary cannot do. If the permitted use is the least useful one, the permission is largely symbolic.
- Verification is undefined. Party documents may contain hidden instructions aimed at AI tools and must be checked before use, but the guidelines do not say what adequate verification looks like. A duty without a standard is hard to comply with and hard to enforce.
- The risk is aimed at the wrong person. Shniger argues the danger comes from users who misunderstand AI, not from sophisticated, screened judges. He would require mandatory training before any use, rather than a public document that discourages use. The guidelines promise training and institutional support, but as an accompaniment, not a precondition.
The pro-innovation concern is not that Quebec drew a line at reasoning. It is that a text which says little about how to use tools well tends to produce unsupervised use. A judge who has never been trained, and who is told only that use is risky, may still reach for a consumer chatbot at 11 p.m. with a deadline looming. A rule that is easy to follow in the abstract, but offers no sanctioned and secure alternative, invites exactly the hallucination episode that triggered it.
How this sits against national guidance
The Canadian Judicial Council issued its own guidelines on AI in Canadian courts on 24 October 2024. They stress judicial independence and ethics, include a preamble on no delegation of judicial decision-making, and acknowledge that translation, grammar checking and legal research tools are already part of routine judicial work. Shniger calls them "quite general" and Quebec's text "more precise and better written". Quebec has therefore moved from a general no-delegation principle to specific barred functions. That is a gain in clarity, and it is also where the tone becomes more restrictive.
Shniger also faults both documents for treating AI as one more tool, when in his words, "It changes how judges work." He contrasts them with guidance from the Action Committee on Modernizing Court Operations, which starts from AI's benefits and helps judges decide whether and how to use it.
A proportionate path
Four changes would keep Quebec's principle and fix its weaknesses:
- Procure and deploy a secure, vetted institutional tool, so that the confidentiality bar does not push judges toward consumer products.
- Make training a precondition of any use, not a companion to it.
- Define verification: what must be checked, against which sources, and how it should be recorded.
- Publish a short list of worked examples of acceptable use, so that "when in doubt, don't" is the exception rather than the rule.
None of this weakens the human-judge principle. It makes the principle operable. Courts that say only what judges may not do will have little to show for it except fewer public disclosures of what judges are in fact doing.