When a British Columbia judge ordered a family lawyer to cover her opponent's costs for filing two cases that ChatGPT had invented, Zhang v. Chen, 2024 BCSC 285 looked like an isolated cautionary tale. It was not. On May 27, 2026, Canadian Lawyer reported that courts and tribunals across Canada have now issued at least 150 more decisions dealing with AI-hallucinated citations and arguments — a count drawn from researcher Damien Charlotin's tracking database, with the vast majority of instances occurring through 2025. What is striking is not the volume but the consistency. Across provinces, courts, and self-represented litigants, Canadian judges have converged on a single principle: the human who signs the document is responsible for what is in it.
The rule the cases keep restating
The sharpest articulation came from the Alberta Court of Appeal. In Reddy v Saroya, 2026 ABCA 20, the court ordered counsel to pay CAD 17,550 plus GST in personal costs after an appellant's factum cited authorities — nearly half of them non-existent — that appeared to have been produced by generative AI. Crucially, the factum had been prepared by a third-party contractor. The court was unmoved: "The use of a third party contractor did not absolve counsel of the need to adequately review work prepared by someone else." The lawyer whose name appears on the filing bears ultimate responsibility, and verifying AI output is now part of ordinary practice management.
That is the throughline of the entire body of rulings. As litigators quoted by Canadian Lawyer put it, the decisions "offer few new insights" precisely because they restate an existing duty rather than inventing a new one. "You can't delegate your professional judgment," Norton Rose Fulbright's Maya Medeiros observed; "lawyers already knew this." The cases are less about artificial intelligence than about an old obligation — to check your sources — applied to a new tool that makes not checking dangerously easy.
The strongest case for a harder line
There is a serious argument for going further than costs awards. Fabricated authorities are not victimless errors: they force opposing counsel to chase phantom cases, consume scarce judicial time, and — at scale — risk seeding the case law itself with fictions if a hallucinated citation slips through unreviewed. A litigant who cannot afford to fact-check an opponent's brief is disadvantaged twice over. Some judges abroad have responded with standing orders requiring lawyers to certify or disclose any AI use, and one can see the appeal of a bright-line rule that removes any ambiguity about what diligence requires.
That concern is real, and the harm is concrete. But it does not follow that courts should ban AI from legal practice or treat every use as presumptively suspect. The Canadian approach has been more proportionate, and better for it.
Why holding the human responsible is the right design
The genius of the Canadian response is that it regulates conduct, not technology. A lawyer who files fabricated cases is sanctioned whether the fabrication came from ChatGPT, a careless articling student, or the lawyer's own imagination. The duty of competence and candour to the court already covered this; the courts simply applied it. That keeps the rule durable as the tools change, and it avoids the trap of writing technology-specific prohibitions that are obsolete by the time they are enforced.
It also preserves the upside. Generative AI genuinely lowers the cost of legal research, drafting, and document review — gains that matter most for small firms and self-represented litigants, the very people a blanket ban or onerous disclosure regime would burden disproportionately. A rule that says "use the tool, but verify the output and own the result" captures the benefit while pricing in the harm. A rule that says "do not use the tool" forfeits the benefit to eliminate a harm that diligence already addresses.
The line judges have drawn around themselves
The accountability principle has a mirror image on the bench. The Canadian Judicial Council's Guidelines for the Use of Artificial Intelligence in Canadian Courts, issued in September 2024, hold that judges "are not permitted to delegate decision-making authority, whether to a law clerk, administrative assistant, or artificial intelligence (AI) computer program, regardless of their capabilities." The Federal Court has gone further still, committing that it "will not use AI, and more specifically automated decision-making tools, in making its judgments and orders, without first engaging in public consultation," and insisting on a "human in the loop" for any AI-assisted task.
This is the same logic pointed inward. Just as a lawyer cannot outsource professional judgment to a model, a judge cannot outsource adjudication to one. The exclusive authority to decide — and the accountability that comes with it — stays with the human. That symmetry is what makes the Canadian framework coherent: responsibility is non-transferable on both sides of the bench.
A model worth exporting
Canada has, almost by accident, produced one of the more sensible templates for governing AI in high-stakes professional settings. It did not pass an emergency statute or stand up a new regulator. It let existing duties — competence, candour, judicial independence — do the work, and let a growing line of decisions make the consequences unmistakable. The CAD 17,550 in Reddy v Saroya is a more effective deterrent than any prohibition, because it tells lawyers exactly what is expected: use the tools if they help, but verify, and never pretend the machine is the one accountable. As other jurisdictions reach for blanket bans and disclosure mandates, the quieter Canadian answer — keep the human responsible — looks like the one most likely to last.