In June, a Law Society Tribunal in Ontario ordered a suspended lawyer to pay $31,150 in costs after his filings turned out to be full of AI-generated citations, some pointing to cases that don't exist and others pointing to real cases that had nothing to do with the argument he was making. The tribunal treated his unchecked use of AI as a significantly aggravating factor, and the cost order was described as the largest of its kind issued by any Canadian court or tribunal to date. That's a striking number on its own, but it's really just the tail end of a much bigger pattern.
According to a tracker maintained by the legal technology site courtready.ca, Canadian courts and tribunals had flagged AI hallucinations or other AI misuse in 239 decisions as of mid-August 2026, and 223 of those involved fictitious citations, hallucinated propositions of law or fabricated quotations. That's up from a small handful in 2024. Most of these cases involve someone representing themselves without a lawyer, though the biggest penalties have landed on lawyers. Sanctions have ranged from a simple warning up to affidavits being struck, motions dismissed, and matters referred to a law society. The dollar figures have been climbing too: before the Ontario order above, the record was $17,550, set in January by the Alberta Court of Appeal in Reddy v. Saroya, payable personally by the lawyer whose factum leaned on seven cases that don't exist.
This isn't only a law firm problem
Court filings just happen to be where AI hallucinations get caught and made public, because a judge is checking the work against a record that everyone involved can see. Every other professional service has the exact same failure mode without the same visibility. An accountant using AI to draft a memo on a tax position, an engineer summarizing a technical standard, a consultant pulling regulatory citations into a report for a municipal client: all of these can hand a client fabricated information delivered with the same confident tone as accurate information, and there's no judge cross-checking any of it before it reaches the client. CPA Canada's quality management standard, known as CSQM 1, already requires accounting firms to document the controls around any technology that affects the quality of an engagement, and AI counts. The expectation isn't that a firm avoids AI. It's that someone can show, after the fact, what was checked and by whom.
The paper trail is the defense
The pattern across these court cases is that the sanction gets worse the less it looks like anyone checked the work. A single wrong citation caught before filing is a non-event. Multiple fabricated cases spread across four separate documents, unnoticed until opposing counsel or a judge flagged them, reads as recklessness, and that's what tribunals keep punishing hardest. For a service business the equivalent habit is easy to describe and easy to skip under deadline pressure: whenever AI output touches a client deliverable, someone verifies the specific facts, figures or citations against a primary source, and that check gets a name and a timestamp attached, even if it's just a one-line note in the file. That record is what turns "we used AI and made a mistake" into "we had a process, and this one slipped through," and that distinction matters enormously to a regulator, an insurer or a court trying to decide whether a failure was an accident or negligence.
Insurers are already asking the question
Speaking on a panel at the National Insurance Conference of Canada, Clyde & Co partner Nathalie David made a point worth sitting with, as reported by Insurance Business Canada: businesses shouldn't wait for AI-specific legislation, because the legal exposure already exists. Contract law, tort law and professional liability rules don't stop applying just because a new tool was involved, and insurers are paying close attention to how the organizations they cover are using AI and whether anyone can explain what their systems actually did.
Applying isn't the same as covering cleanly, though. By 2026, several major carriers had added explicit AI language to their errors and omissions policies rather than leaving the question open, and underwriters are increasingly asking firms directly whether they use AI, whether someone reviews its output, and whether that review is documented, before they'll price a policy. A firm with a written verification step can answer that question in one sentence. A firm without one is negotiating from a weaker position, and may only find that out once a claim is already on the table.
None of this requires a formal AI policy binder or a compliance department. It requires deciding, in writing, who checks AI-generated facts before they leave the building, and keeping a light record that it happened. That's the same paper-trail advice compliance people have been giving for other technology risks for years. AI just made it urgent again, and the hundreds of flagged decisions on that tracker are what happens when nobody wrote it down.