A Certificate, Not a Ban
On June 2, 2026, Chief Justice Suzanne Duncan of the Supreme Court of Yukon issued General Practice Direction 35, titled Conditions on Use of Generative AI in Written and Oral Representations. It requires all counsel and self-represented litigants to file a signed certificate confirming the accuracy of every legal authority cited in written submissions and oral arguments. Non-compliance may result in costs assessments — including personal costs against counsel, not merely the parties they represent. The direction explicitly identifies the target behaviour: AI hallucination, defined as generative AI "hallucinating" case law that does not exist.
This is not a ban on AI in legal work. Chief Justice Duncan's direction acknowledged that generative AI "is a tool that may facilitate some of the work of counsel and litigants" and encouraged "discussion and collaboration with stakeholders." What it does is close a doctrinal gap that courts across Canada have watched widen since 2023: the gap between a lawyer's existing professional obligation to verify every citation and the practical temptation to trust AI-generated research at face value.
A Documented Canadian Problem
The direction arrives against a backdrop of growing documented harm. The courtready.ca sanctions database, which tracks AI hallucination proceedings in Canadian courts and tribunals, records 195 proceedings across 53 courts as of June 23, 2026. Of those, 164 — roughly 84 percent — involved self-represented litigants without counsel. The largest costs award on record in Canada exceeds $31,000. The database notes that the most serious sanctions, including the highest costs orders, have come in 2026.
High-profile cases have sharpened the stakes. In Hussein v. Canada (2025 FC 1060), a Federal Court judge discovered that counsel had submitted multiple non-existent case citations generated by the AI tool Visto.ai. Four separate court directions were required before counsel acknowledged the AI's role. Associate Judge Catherine Moore awarded special costs personally against counsel — not the client — and wrote that while AI use "is a perfectly valid tool for counsel to use," its output "must be verified by a human" and its use "must be declared." In Zhang v. Chen (2024 BCSC 285), BC Supreme Court similarly found counsel personally liable for costs after ChatGPT fabricated the citations that appeared in counsel's filings. Ko v. Li (2025 ONSC 2766) escalated further, ordering a contempt hearing — resolved only through mandatory professional development attendance.
Canada's Patchwork Becomes a Framework
Yukon's June 2026 direction is the most recent layer of what is becoming a national framework by accumulation rather than by design. Manitoba's Court of King's Bench issued a disclosure requirement on June 23, 2023. Yukon itself had General Practice Direction 29 from June 2023, requiring disclosure of AI use and the specific tools employed. Nova Scotia's Provincial Court issued guidance in October 2023. The Federal Court published a practice notice in May 2024. Ontario amended its Rules of Civil Procedure through O. Reg 384/24, effective December 1, 2024, requiring parties to certify cited legal authorities are "authentic." Tribunals Ontario followed with its own guidance in April 2025.
What distinguishes the June 2026 Yukon direction is its enforcement posture. Most prior directions required disclosure or encouraged verification. General Practice Direction 35 requires a signed certificate — a formal attestation whose absence triggers enforceable sanctions. That is a meaningfully higher accountability threshold, and it shifts the burden from post-hoc discovery of mistakes to pre-filing verification by design.
Steelmanning the Case for Strict Rules
The regulatory logic here is compelling, and courts deserve credit for articulating it. Legal citation is not a domain where near-miss accuracy is tolerable. When a lawyer cites a non-existent precedent, the opposing party may not catch it, the judge may rely on the fabricated holding, and a real person's rights may be affected. Existing professional responsibility frameworks already require verification; the certificate requirement simply makes that obligation legible, formal, and sanctionable. From that view, Yukon's direction is not creating a new duty — it is making an old one enforceable with consequences proportionate to the professional gravity of a court submission.
Proportionate, With One Calibration Risk
What makes this direction proportionate is what it leaves alone. It does not dictate which AI tools are permissible. It does not require prior court approval for AI-assisted research. It does not mandate disclosure of every AI interaction — only that whatever authorities are ultimately cited have been independently verified. The accountability falls exactly where professional responsibility already places it: with the person who signs the filing.
The calibration risk lies in the self-represented cohort. The courtready.ca data showing 164 of 195 Canadian AI hallucination proceedings involve unrepresented parties underscores that the certificate model must be explained accessibly, or it becomes a trap for litigants who do not understand what they are signing or why CanLII verification is required. Chief Justice Duncan's call for "discussion and collaboration with stakeholders" is worth taking seriously here — the framework will need plain-language guidance for self-represented users if it is to deter harm rather than just punish it after the fact.
The Enforcement Model Worth Watching
Yukon's approach demonstrates that courts do not need to await parliamentary AI legislation. Courts hold inherent jurisdiction over their own proceedings, and practice directions are a nimble tool. The certificate model — assign clear accountability, specify consequences, leave tool selection to the practitioner — has the right architecture for AI governance in professional settings: it is outcome-focused rather than technology-prescriptive, which means it will not become obsolete as the underlying tools evolve.
If Yukon's early experience shows that signed certificate requirements reduce hallucination incidents without chilling legitimate AI-assisted research, expect other provincial superior courts to follow. For Canadian lawyers, the immediate implication is simple: before any authority is cited, check it in a reliable database, and be prepared to certify that you did.