‘A large language model does not appreciate nuance or exercise judgment or use a moral compass when writing’
A suspended Ontario lawyer must pay the province's legal regulator $31,150 after filing motion materials drafted by generative artificial intelligence (GenAI) that cited fabricated and irrelevant case law, the Law Society Tribunal has ruled.
In Mazaheri v Law Society of Ontario, a three-member panel found that Shahryar Mazaheri's irresponsible use of AI was a significantly aggravating factor and ordered him to pay full costs by June 26 this year, with four-per-cent interest on any overdue amount.
It is the first time a party has submitted "hallucinated" authorities to the tribunal.
Factum, affidavits done by AI
The panel had suspended Mazaheri's licence on an interlocutory basis in November 2024 amid allegations he was involved in fraudulent mortgage transactions. He later brought two motions — one to cancel the suspension, another to exclude the regulator's evidence and have the panel recuse itself. Both failed.
The tribunal declined his request to defer costs to the future panel that will hear the misconduct allegations, saying it was better placed to assess motions it had heard. It stressed that it made no finding of misconduct, noting a separate conduct application remains pending.
Mazaheri filed a factum, supplementary factum and two affidavits – all produced with generative AI. The materials cited tribunal decisions that did not exist, referenced real cases that did not support his arguments, and misused procedural rules as substantive law, according to the panel.
The tribunal said not one proposition was backed by a reliable authority, calling the result "gibberish."
After being flagged, Mazaheri admitted using the tool without checking it, apologised and undertook not to use it again. In a written statement narrating how the errors arose, he said:
"These errors are entirely my responsibility. They arose for [sic] my over-reliance on generative artificial intelligence tools (in particular Grok) to assist me in researching and drafting the documents while I was trying to manage filing these materials on my own… I did not intend to mislead the tribunal. I believed in good faith that the AI tools would produce accurate citations and reasoning. I now see that I failed to verify the output with sufficient care."
HR professionals risk eroding employee trust, accelerating attrition, and exposing their organizations to legal and reputational damage as workplace AI adoption outpaces ethics and compliance policies, according to one expert.
Why did AI tools fail?
The Law Society Tribunal panel devoted much of its decision to why AI research is unreliable — reasoning that reaches beyond law. A large language model, it said, calculates statistical relationships between words and assembles authoritative-sounding text, but does not reason or verify.
"But unlike a human, a large language model does not appreciate nuance or exercise judgment or use a moral compass when writing. It just does calculations, and then puts words in the order dictated by statistical probabilities. The calculations are impressive – large volumes of authoritative-sounding text are generated in seconds or minutes. But they are just calculations. And the product only reads or sounds like text written by humans. It isn't the same."
Mazaheri said his over-reliance arose because he could not retain a lawyer. The panel rejected that as "a non-explanation," reasoning that he remained responsible regardless of who assisted him — a principle that applies to any manager signing off on AI-assisted work.
The tribunal did not treat AI itself as the problem, only its unsupervised use, noting that users must still review and correct output manually. It said it would have awarded most costs in any event, but that the false authorities "justifies awarding full costs against the respondent."
Fabricated citations are rising in Canadian proceedings, the tribunal added, counting 132 cases on CanLII — seven in 2024, 86 in 2025 and 39 in early 2026. In 24 of them, a represented party, not a self-represented one, submitted the false citations.
AI “workslop” – defined as “AI-generated work content that masquerades as good work” – is now “ubiquitous," according to Fenwick McKelvey, associate professor of communication studies at Concordia University and co‑director of the Applied AI Institute.
According to Tom Macintosh Zheng, a Toronto-based lawyer and co-founder of Courtready.ca, that $31,150 fine against Mazaheri sets the record “by far” for AI-related cost impositions. “Until this case came along, the largest penalty issued by a Canadian court was in a case called Reddy v. Saroya, which is a 2026 Alberta Court of Appeal decision. It was $17,550,” he said, according to Law Times.
In an email to CHRR, Zheng said that we now have 168 decisions across 51 courts or tribunals where someone was "caught submitting fabricated case law".
"Of the 168 decisions, 81% were self-represented litigants, and 19% were represented parties," he added. "We are also seeing a new frontier of cases involving other types of AI-misuse, such as fabricated quotes from real decisions, hallucinated propositions of law, and AI-generated evidence."
