A factum was filed in an Ontario estates matter citing four authorities. Three of them did not exist. The fourth was real, and said the opposite of what it was cited for.
In July, the Law Society Tribunal suspended the lawyer who filed it for six months and ordered her to pay $10,000 in costs; believed to be the first time a Canadian legal regulator has suspended a licence in a case involving AI-generated case law. Six months is close to the most severe sanction available short of revocation.
Law Society of Ontario v. Lee, 2026 ONLSTH 136, has been reported across the profession, and in the UK and Australia, as the moment Canadian regulators drew a line on artificial intelligence. This framing is understandable and mostly wrong. The tribunal said, in terms, that using AI to help prepare a factum is not professional misconduct. The suspension was not for using ChatGPT. It was for what happened after.
For any firm currently deciding how, or whether to let AI into its workflow, that distinction is the entire lesson. If you read it as a technology case, you might draw the wrong conclusion about your own exposure.
What Happened
The underlying matter was an estates dispute. The lawyer, a sole practitioner of roughly thirty years’ call, filed a factum prepared with ChatGPT. It cited a case as authority for removing an estate trustee; in the actual decision, the court had declined to remove the trustees.
Hon. Justice Fred Myers of the Superior Court asked for copies of the authorities. She could not produce them. The judge went looking himself and found that some of the cited decisions could not be located in any database, and that others were irrelevant or contradicted the argument they were offered to support.
When asked directly in open court whether AI had been used, the lawyer said no. She later told both the court and the Law Society that staff in her office had used ChatGPT without her knowledge, and then eventually acknowledged that she had not employed an associate for years.
The tribunal made three findings: she failed to serve her client, she misled the court on multiple occasions, and she failed to act with honour and integrity toward both the court and the regulator.
A centralized matter management system allows legal teams to:
A centralized matter management system allows legal teams to:
- Store all case-related information in one location
- Access files instantly from anywhere
- Track important milestones and deadlines
- Collaborate efficiently across departments
- Maintain complete case histories
The line the tribunal actually drew
The reasons are careful, and worth reading in full rather than in headline form. The tribunal stated plainly that it is not misconduct for a lawyer to use AI to assist in preparing a factum. The failure was not the tool, but was filing a document without verifying what was in it and then, far more seriously, being dishonest about it.
The sanction reflects that weighting. The tribunal considered whether six months should have been longer, and settled where it did partly because it accepted the lawyer’s remorse as genuine and the conduct as out of character, and because she was already facing significant consequences, including contempt proceedings later this year.
This is where the “first AI suspension” label needs a qualifier. Counsel tracking these cases across Canada have cautioned against reading the decision as establishing that AI misuse equals six months. That reading skips the dishonesty which is what the tribunal itself identified as the aggravating factor. The case is a first. It is not yet a benchmark, and firms treating it as a fixed tariff for hallucinated citations are reading a different decision than the one that was written.
There is a useful comparison in Mazaheri v. Law Society of Ontario, 2026 ONLSTH 112, decided weeks earlier. There, a self-represented licensee filed multiple AI-generated documents citing decisions that did not exist and real cases that did not support his position. He admitted it, apologized, and committed to stopping. He was ordered to pay just over $31,000 in costs a record for AI-related cost sanctions in Canada, but the tribunal made no misconduct finding on that motion.
Same underlying failure. Very different response. The variable was candour.
The Volume Problem
This is not an isolated pair of cases. Counsel tracking fabricated citations on CanLII have reported roughly seven such decisions in 2024, in the high eighties in 2025, and dozens more in the first half of 2026 alone. Reported counts differ somewhat by source and date, but the trajectory is not in dispute.
The more uncomfortable number is the breakdown: while most involve self-represented litigants, a meaningful minority reported at close to one in five involve represented parties. That is lawyers filing unverified AI output.
What firms should take from this
Six practical points:
1. Verification is the control, not abstinence. Firms that ban AI outright are managing the risk on paper while it walks in through personal devices. A policy that assumes use and requires verification is more honest and more effective than one that assumes non-use.
2. Every citation gets checked against the source. Not the summary, not the hyperlink but the actual decision, in an actual database, read for whether it holds what it is cited for. In Lee, the fourth case existed and still sank the factum.
3. Name who is accountable before it matters. The lawyer who signs the document owns its contents. Making that explicit in a firm policy avoids the improvisation that followed here.
4. Supervision duties are unchanged. Students, clerks and assistants using AI is a supervision question the profession already has rules for. The tool is new; the obligation is not.
5. Build a response plan for the moment an error surfaces. Almost every escalation in these cases traces to what happened after the problem appeared. Correcting the record immediately is unpleasant. It is also, on this record, the cheapest available option by a wide margin.
6. Log it. Which tool, for which task, who reviewed the output. If a court or the LSO ever asks, the answer should be retrievable rather than reconstructed.
The Uncomfortable Read
The profession’s instinct after Leehas been to treat it as a warning about technology. It reads better as a warning about process. A thirty-year sole practitioner filed something she had not checked, was asked about it, and gave an answer she could not sustain.
Firms adopting AI well and firms adopting it badly are not distinguished by whether they use it. They are distinguished by whether anything reads the output before it leaves the building.
This post is general commentary on a published tribunal decision, not legal advice. Reasons in Law Society of Ontario v. Lee, 2026 ONLSTH 136 are available on CanLII.