Mazaheri v Law Society of Ontario, 2026 ONLSTH 112
Shahryar Mazaheri is a lawyer whose licence was suspended on an interlocutory basis in November 2024. He brought a motion to cancel or vary that suspension. When he filed his supporting materials, those materials were riddled with citations to cases that do not exist, citations to real cases that stood for propositions entirely unrelated to the points he was making, and repeated references to the Tribunal’s Rules of Practice and Procedure as though they were sources of substantive law.
The Tribunal discovered the problem, convened a case management conference, and gave Mazaheri an opportunity to explain himself. He admitted he had used generative AI, specifically Grok, and had failed to check the output before filing it. He apologized and was permitted to refile.
The motions panel ultimately rejected both of his motions, awarded the Law Society its full costs of $31,150, and treated the AI hallucinations as a significant aggravating factor in that award.
The costs outcome matters and we return to it below. But the most valuable part of this decision for practitioners is the panel’s explanation of why large language models are structurally incapable of doing what lawyers do, and why submitting their unverified output to a tribunal is not just careless. It is a failure of professional responsibility.
Why a Large Language Model Cannot Think Like a Lawyer
The panel’s explanation is worth reading carefully because it actually tackled the issue of why and where AI will fail if left to its own devices to do legal work.
As Tribunal explained, a large language model is trained on vast quantities of text. It identifies statistical relationships between words and phrases and then calculates the order in which to arrange them so that the output sounds coherent and authoritative. The result reads like something a human wrote, however there are are several substantial why there are serious deficencies in AI work.
The panel named several issues that make large language models dangerous in a legal context. For purposes of this article, I focus on three that I believe are the most salient for practitioners working in regulated settings.
First, a large language model lacks judgment. It cannot verify whether a case it cites actually supports the argument it is being asked to make. It cannot critically assess a factual matrix against the authorities it generates. It produces text that sounds right without any mechanism for checking whether it is right.
Second, a large language model does not know the limits of its knowledge. The panel put it directly: an attribute of human judgment is knowing when to admit that you do not know something. A large language model is strongly predisposed to giving an answer, any answer, rather than admitting ignorance. That predisposition is structurally baked in. It cannot be turned off by asking the tool to be more careful.
Third, a large language model is not aware when it is inventing something. A human who fabricates a citation knows they are doing it. A large language model does not. It is calculating probabilities and assembling text that has a high statistical likelihood of sounding coherent. The fabrication is not deliberate. It is inevitable.
These three deficits explain why Mazaheri’s materials were not just wrong. They were confidently, authoritatively, structurally wrong in a way that required the panel to stop everything, document the extent of the problem, convene a case management conference, and get the proceedings back on track. That is what unverified AI output does in a legal proceeding. It does not just fail. It fails in a way that generates significant collateral damage.
The Responsibility Does Not Disappear Because the Tool Generated the Content
Mazaheri’s explanation was that he was managing the filings on his own and could not retain a lawyer to assist him. The panel dismissed this as a non-explanation. It would not have mattered if he had an army of lawyers working for him. He was responsible for what he submitted. The tool that generated the content does not share that responsibility.
The panel quoted Myers J. of the Ontario Superior Court on this point, summarizing a lawyer’s obligations in terms that apply equally to any licensee submitting materials to a tribunal: it is the lawyer’s duty to ensure human review of materials prepared by non-human technology. It is the lawyer’s most fundamental duty not to mislead the court. And at its barest minimum, it is the lawyer’s duty not to submit case authorities that do not exist or that stand for the opposite of the lawyer’s submission.
None of those duties are discharged by the fact that a tool produced the content. They are discharged by the person who submits it.
The Costs
The panel made clear it would have awarded most of the Law Society’s costs regardless of the AI issue. Both motions failed. The litigation was unnecessarily prolonged by the respondent’s own procedural choices. The Law Society was entitled to oppose both motions given the seriousness of the underlying allegations, which included large scale mortgage fraud connected to a double murder and a suicide.
The AI hallucinations were an additional and significantly aggravating factor. The panel’s approach is consistent with the direction courts across Canada have taken: full costs, solicitor-client costs, and in some cases costs against counsel personally. The message from every court and tribunal that has addressed this issue is the same. Irresponsible use of AI in legal proceedings is not treated as a technical error. It is treated as a failure of professional responsibility with financial consequences to match.
What This Means for Licensed Professionals
The panel was explicit that AI tools are not the problem. Irresponsible use of them is. A lawyer who understands the law, develops prompt engineering skills, and verifies the output against primary sources is using the tool appropriately. A lawyer or licensee who accepts the output uncritically and files it is not.
The numbers regarding hallucinated citations the panel cited are striking. Hallucinated citations were identified in Canadian cases 132 times between 2024 and the first quarter of 2026. In 24 of those 132 cases, the party submitting the hallucinated material was represented by counsel. This is not a self-represented litigant problem. It is a profession-wide problem.
And the response from courts and tribunals is consistent: full costs, solicitor-client costs, costs against counsel personally. The message is not subtle.
If you are preparing materials for a regulatory proceeding, a discipline hearing, or any tribunal process, and you are using AI to assist with legal research or drafting, the verification obligation is not optional and it is not delegable. Whatever the tool produces, you own.
FAQ
1. I used AI to help me draft materials for a regulatory proceeding. Am I responsible if the materials contain errors or hallucinated citations?
Absolutely. The Mazaheri decision makes clear that whether you drafted the materials yourself, had a colleague draft them, or used a generative AI tool, you are the person who submitted them and you bear full responsibility for their accuracy. The Tribunal rejected Mazaheri’s explanation that he was managing the filing on his own as a non-explanation. Inability to retain assistance does not discharge the obligation to verify what you submit.
2. What is an AI hallucination and why is it a problem in legal proceedings?
A hallucination occurs when a large language model generates content that is factually wrong but presented with confidence. In a legal context this most commonly means citations to cases that do not exist, citations to real cases that do not stand for the propositions attributed to them, or references to rules and statutes applied in ways that have nothing to do with their actual content. The problem is not just that the content is wrong. It is that it is wrong in a way that is designed to sound authoritative, which means it can mislead opposing counsel, adjudicators, and the process itself before anyone catches it.
3. What are the cost consequences of submitting hallucinated AI materials to a tribunal?
The Mazaheri panel awarded the Law Society its full costs of $31,150 and treated the AI hallucinations as a significantly aggravating factor. This is consistent with the approach taken by courts across Canada, where irresponsible use of AI has attracted solicitor-client costs, full indemnity costs, and in some cases personal costs orders against counsel. The common thread in every decision is that submitting unverified AI output is not treated as a technical error. It is treated as a failure of professional responsibility with financial consequences at the higher end of the spectrum.
4. I am facing a regulatory proceeding and I cannot afford a lawyer. Can I use AI to help me represent myself?
You can use AI as a drafting and organizational tool, but you cannot rely on it for legal research or for anything else without verifying every output against primary sources. The Mazaheri decision arose in exactly that context: a self-represented licensee who used AI because he could not retain counsel. The panel had no sympathy for that explanation. If you are facing a regulatory proceeding and representation is a concern, there are options worth exploring, including limited scope retainers where a lawyer assists with specific parts of the matter rather than the whole file. That is a conversation worth having before you file anything.
Anna Tamir is a regulatory defence lawyer and the principal of Tamir Litigation Law Firm in Richmond Hill, Ontario. Her practice focuses on defending licensed professionals before Ontario’s regulatory bodies, including the Law Society of Ontario, the CPSO, the CICC, and others. She can be reached at info@tamirlitigation.com or 416 499 1676, or at tamirlitigation.com.
This commentary is for informational purposes only and does not constitute legal advice.
Ready for the URL to prep the schema.