Beyond Fake Cases: The Other Ways AI Is Going Wrong in Canadian Courts

Earlier this year, a motion at the Ontario Superior Court paused while everyone in the room went looking for a quotation. The factum on one side quoted a decision of the Court of Appeal, and opposing counsel could not find the quoted words anywhere in that decision. He suspected the factum had been drafted with AI. The judge called a short recess, so the party who filed it could go and find the passage.[1]

The citation was correct, and anyone who looked it up would have found exactly the case named. In this instance, AI produced a fake quotation from a real case. A legal database would have cleared the citation without trouble, because the citation was never the problem.

For two years now, the story of generative AI in Canadian courts has been a story about fake cases. We have catalogued the problem in more than 166 published decisions. However, behind them sits a second body of AI misuse that the defences built for fake cases do not reach. We have now found 28 Canadian decisions in which a court or tribunal flagged AI misuse that was something other than a fictitious citation.[2] In general, they come in three kinds.

Propositions with no source

The first kind is a hallucinated proposition of law. The tool states a legal rule with confidence, and either gives no citation at all or points to a real provision that says nothing of the kind. In one strata dispute before British Columbia’s Civil Resolution Tribunal, a self-represented litigant built his argument on six sections of the Strata Property Act. While the sections are real, none of them said what the party claimed they said, and the tribunal treated the submissions as likely AI hallucinations.[3] A near-identical thing happened in another strata matter weeks earlier.[4]

Fake quotes from real cases

The second kind is the one that stopped the motion described above. The case is real and correctly cited, but the words attributed to it are invented. In one Ontario proceeding, a party’s reply factum quoted a passage it ascribed to Justice Rothstein, writing for the Supreme Court of Canada in Sattva. The court found that nothing resembling the quotation appears anywhere in the decision.[5] In a tribunal matter, a party quoted a block of text from paragraph 150 of a decision that only has 124 paragraphs.[6]

The new frontier: evidence

The third kind is the newest, and it is the reason this is worth writing about now. Litigants have begun using these tools not to find the law, but to create evidence.

For example, a party in BC supported his case with material generated by ChatGPT and Claude, which the tribunal disregarded.[7] In another, an applicant asked a chatbot to describe the signs of a horse that needs dental work, then asked it to review a clinic’s records and summarize the veterinarian’s findings, and submitted the output as proof.[8] Before Quebec’s labour tribunal, a worker put forward a large body of AI-generated research to establish that his shoulder injury had been caused by his job, and the tribunal dismissed the material in its entirety.[9]

And notice where these cases are landing. Of the 28 decisions, 25 involved self-represented litigants, and 20 came from tribunals rather than courts. Almost none drew a consequence heavier than a warning, with the highest costs award in the set being $910.[10] Tribunals are the busiest and least formal corners of the system, where the rules of evidence are relaxed by design, and where an adjudicator working through a long list of cases will sometimes decide on the arguments that hold and set the rest aside. It is the softest target in the system, and it is where the harder problem is arriving first.

Where the rules stop

None of this is what the current rules were built for. Most practice directions issued so far speak to the same hazard, that AI invents fictitious citations, and tell parties to verify their case law against a trusted database like CanLII.[11] The whole apparatus appears to stop at the citation, which is exactly where the new frontier begins. There is no CanLII for facts, and almost nothing in the rulebooks points at them yet.

So, the response must move onto the evidence layer. Right now, when a tribunal confronts AI-generated evidence, it improvises. It reaches for a thin line of earlier decisions that declined to rely on such material because of its inherent unreliability, and it decides the question case by case.[12] That instinct is sound, but it should not have to be rebuilt every time. British Columbia’s Civil Resolution Tribunal has shown what the alternative looks like. As of February 1, 2026, its rules forbid participants from submitting fabricated evidence, including evidence created or altered by an AI tool.[13] A rule like that is the first half of the answer. The second half is a default the practice directions could adopt tomorrow: AI-generated material offered as evidence carries no weight until the party putting it forward can show where it came from and why it can be trusted.

This will not catch everything, and we have to be honest about that. A standard like this works on the litigant who discloses, or whose opponent objects. It does nothing about the determined party who launders an AI-generated document into their record and says nothing. But it is a starting point to address this new frontier.

________

Tom Macintosh Zheng is a Toronto-based former commercial litigator and the co-founder of Courtready.ca. Courtready builds practical tools for Canadians who litigate and conducts original research on access to justice issues. He maintains a live database of Canadian decisions flagging AI-fabricated case law and other AI misuse, updated daily, at courtready.ca/fictitious-citations-in-canadian-courts. He is the 2026 recipient of the Ontario Bar Association Foundation Award.

________

[1]RSR Road Surface Recycling v Bonnechere Excavating et al., 2026 ONSC 698 at paras 34, 39.

[2]Figures in this article are drawn from the author’s database of Canadian decisions flagging AI misuse, current to June 2026 and updated daily: courtready.ca/fictitious-citations-in-canadian-courts.

[3]Wyszynski v. The Owners, Strata Plan BCS1721, 2026 BCCRT 851 at para 10.

[4]Chen v. The Owners, Strata Plan LMS3094, 2026 BCCRT 823 at para 11.

[5]Kapahi Real Estate Inc. v. Elite Real Estate Club of Toronto Inc., 2026 ONSC 1438 at paras 24–25.

[6]Re X Corp., 2025 BCCRT 1228 at para 45.

[7]Lee v. AutoCanada MR Motors GP Inc., 2026 BCCRT 854 at para 8.

[8]Iida v. Meadow Lane Equine Clinic Ltd, 2026 BCCRT 626 at paras 25–26.

[9]Côté et Placements JF Desgagné inc., 2026 QCTAT 1311 at paras 21–23.

[10]Ren v. Area 09, 2025 BCPAAB 20253891, at para 33.

[11]See, e.g., Tribunals Ontario, Practice Direction on the Use of Artificial Intelligence (AI) in Tribunal Proceedings.

[12]See, e.g., Bahia v. ICBC, 2025 BCCRT 453 at para 14.

[13]Civil Resolution Tribunal, Standard Rules, r 8.3(9) (effective February 1, 2026).