A factum made almost entirely of bad authority

The discipline order against Ontario lawyer Mary Hyun-Sook Lee, who also practises under the name Jisuh Lee, arrived on July 16, 2026. The Law Society Tribunal found that she had failed to serve her client by relying on a factum containing only nonexistent or irrelevant case law generated by an AI tool. It also found that she deliberately misled the court when discussing the AI-drafted filing and was later dishonest with the Law Society about her use of AI.

The penalty was a six-month licence suspension beginning July 17 and CAD 10,000 in costs. Overdue installments accrue interest at 4%. The order is short and does not publish the tribunal's full reasoning, which matters when assigning weight among the three findings. The suspension was not described as a tariff for merely opening ChatGPT. It followed a failed filing, a false explanation to the court, and dishonesty during the regulator's investigation.

The filing began in a family-law proceeding. In Ko v. Li, Justice Frederick Myers reviewed a factum containing authorities that could not support the propositions attributed to them. Some cases did not exist. Others were real but irrelevant. The document had the appearance of legal research without the underlying law.

A factum is the written argument a Canadian lawyer files to explain the relevant facts, issues, and authorities. The authorities are not garnish. They tell the court why the requested result follows from existing law. If every useful citation is invented or unrelated, the court has not received a weak argument. It has received a counterfeit research trail.

The first explanation blamed staff

When the court asked for an explanation in May 2025, Lee said staff had partly used ChatGPT to prepare the filing and that the false authorities were model hallucinations. She apologized, replaced the factum, and undertook additional training. At that stage, the account located the tool failure inside a delegated drafting process that Lee had failed to supervise.

That account did not survive.

In a September 30, 2025 letter reproduced in a later review of Canadian court cases involving improper AI use, Lee told the court that she had prepared the factum alone. She acknowledged personally using ChatGPT for research and drafting, failing to verify its citations, and falsely implying that staff had been involved. She attributed the misrepresentation to fear of the consequences and embarrassment.

The correction fixed the authorship record and opened a more serious problem. The court had already been examining whether the original filing could amount to contempt. Lee's first response to that inquiry supplied another allegedly contemptuous act: a deliberate false statement to the judge about who did the work.

According to Law Times, a separate criminal-contempt proceeding remained active when the tribunal imposed discipline. The professional order and the contempt process are distinct. The tribunal made misconduct findings and suspended Lee's licence; it did not resolve the criminal-contempt matter.

Verification was the missing legal work

ChatGPT can generate case names, citations, summaries, and quotations that resemble legal research. It cannot make those authorities real. Every citation still has to be located in a reliable database, read in context, and checked against the proposition in the filing. That work is not an optional quality-control pass after the AI finishes. It is the legal research.

Lee's September account said she failed to verify the citations independently. The tribunal's finding was broader: the factum contained only nonexistent or irrelevant AI-generated authority. A single database check would have exposed a nonexistent case. Reading the real cases would have exposed the irrelevant ones. The normal workflow failed at both gates.

This is also why blaming the model never settles a court's concern. ChatGPT did not sign the factum, owe duties to the client, or answer the judge's questions. A lawyer can use drafting software, search tools, junior staff, or a model, but the filed document still carries the lawyer's professional representation that its authorities have been checked.

The later dishonesty finding made the incident worse in a way no improved prompt could repair. A hallucinated citation begins as unreliable output. It becomes professional misconduct when a lawyer files it without verification. A false story about its origin is a separate human decision made after the software is out of the loop.

The order is narrower than the headline

Law Times reported that the case may be Canada's first lawyer suspension partly connected to reliance on AI-generated case law. The qualifier matters. The tribunal's brief order does not isolate how much of the six-month penalty came from the defective filing, misleading the court, or dishonesty with the regulator. Until full reasons are published, claims that AI use alone produced a six-month suspension go beyond the available record.

What the order does establish is substantial enough. A lawyer filed a factum whose legal support was nonexistent or irrelevant, gave the court a false account that shifted responsibility to staff, later admitted personal ChatGPT use, and was suspended after the regulator found three forms of professional misconduct.

The model supplied fake authority. The professional process was supposed to prevent that text from becoming a court filing. Then the disciplinary process asked who was responsible, and the first answer was false. By July 2026, the expensive part of this episode was no longer the hallucination. It was everything a licensed professional did with it.