4 Quebec Courts Restrict Generative AI in 10-Page Rules

Quebec's judiciary has delivered a precedent-setting operational framework for artificial intelligence in adjudication. The 10-page guidance draws a hard line between AI-assisted drafting and human judicial reasoning, while stating that no approved, secure institutional tool currently exists. For legal operations, compliance, and RegTech teams, the document effectively defines the product and governance requirements that any future court-approved AI system will need to meet.

On September 5, 2026, Quebec's four principal courts — the Court of Appeal, the Superior Court, the Court of Quebec, and the province's municipal courts — jointly adopted a 10-page set of guidelines governing judges' use of generative artificial intelligence. The framework is unambiguous on the central point: while generative AI may assist with ancillary tasks, it cannot replace judicial reasoning, the assessment of evidence, or deliberation. Judging remains the exclusive responsibility of judges. The guidelines state that “Generative AI, however sophisticated, possesses neither judgment nor conscience and is not subject to any form of accountability,” and further warn it cannot understand the human, social, and legal context specific to each case.

“

In March 2026, La Presse reported that a Quebec court decision contained references to case law that did not exist, raising the possibility that generative AI had been used in preparing the decision.

The allowed uses are deliberately narrow. Judges may use AI to correct, revise, rephrase, or translate text they have already written, handle certain administrative work, or locate information within specified sources. The guidelines explicitly prohibit relying on AI as a source of legal authority or reasoning, assessing evidence, or making decisions. This creates a practical perimeter: AI is a drafting and retrieval aid, not a deliberative participant. The distinction is significant for court operations because it keeps the core adjudicative function human while acknowledging efficiency gains in routine documentation and translation, especially in a bilingual jurisdiction like Quebec.

The rules did not emerge in a vacuum. In March 2026, La Presse reported that a Quebec court decision contained references to case law that did not exist, raising the possibility that generative AI had been used in preparing the decision. That incident, involving so-called hallucinated citations, created urgency within the judiciary. Although the courts did not confirm the earlier use, the timing of the September guidelines strongly indicates that the phantom-citation episode served as a catalyst. It also aligns with broader awareness across the legal sector that large language models can produce fluent but fabricated legal authorities, a risk that is particularly acute in adjudication where incorrect citations can undermine the integrity of decisions.

The guidelines also address institutional readiness and transparency. The courts disclosed that there is currently no approved, secure institutional generative AI tool available to judges. Existing commercial tools, they note, rely on algorithms whose workings remain largely opaque. This is more than a technical caveat; it signals that the Quebec judiciary will not sanction widespread adoption until security, confidentiality, and explainability requirements are met. For court administrators and legal technology vendors, the document effectively defines the product requirements for any future approved tool: it must be secure, institutionally controlled, auditable, and limited to non-deliberative support functions.

The document also offers a normative statement about the nature of judging. It says “Judging cannot be reduced to a technical exercise” and describes adjudication as an intrinsically human function based on deliberation, consideration of context, respect for dignity, fairness, and moral responsibility. This language matters beyond Quebec because it provides a rationale that other courts and regulators can adapt. It rejects the premise that better training data or more sophisticated models alone will make AI fit for judicial reasoning. Instead, it treats human judgment, conscience, and accountability as non-delegable features of the judicial role.