Eleven courts. Nine law societies. Zero bans. The volume of Canadian guidance on generative AI looks alarming until you notice what it is actually asking for — not prohibition, but discipline around five specific, manageable risks.
Key takeaways
“Since ChatGPT entered the legal lexicon, eleven Canadian courts, nine law societies, two professional liability insurers, one provincial government, the Canadian Judicial Council, Canadian Bar Association and the College of Patent Agents and Trademark Agents have all issued guidance documents regarding the use of Generative AI,” and “none have attempted to ban its use.” (LSA, Gen AI Rules of Engagement) That combination — this much regulatory attention, and zero prohibitions — is itself the finding worth explaining. What are they actually worried about, if not the technology’s existence?
Every court notice surveyed for this hub names the same failure mode by a different route. The Federal Court defines “hallucination” explicitly. (Federal Court, Notice to the Parties and the Profession, May 7 2024) Ontario names it as the mechanism behind its sanctions regime. (Ontario SCJ, Consolidated Civil Practice Direction) Alberta’s notice does not use the word, but its entire Reliance clause exists to prevent exactly that outcome. None of the three treats AI as unreliable by nature — each treats it as a tool whose output has to be checked, the same way a junior researcher’s memo would be checked before it goes in front of a judge.
A quieter, more structural worry runs through the Law Society of British Columbia’s guidance: it is not only that a lawyer might carelessly disclose a client’s facts, but that the tool itself might retain them. “As generative AI tools (especially those powered by LLMs) rely on previously inputted information to learn how to produce more accurate answers in the future, the generative AI tool may reuse the confidential information you supply for other purposes over which you may have no control.” (LSBC, Guidance on Professional Responsibility and Generative AI) Consider what that means concretely: a lawyer pastes a client’s draft settlement terms into a public chatbot to tighten the wording. The immediate act — typing it in — feels no different from using a word processor. The guidance’s point is that it is not the same act, because the recipient on the other end may learn from it in a way a word processor never does.
The confidentiality worry and the competence worry are linked. LSBC ties competence directly to the Code of Professional Conduct: “before using generative AI in your client work, it is prudent to seek out more information and training on how to use the tool,” because Code rule 3.1-2 requires performing legal services “to the standard of a competent lawyer.” (LSBC, Guidance on Professional Responsibility and Generative AI) A lawyer who does not understand how a tool retains, generates or fabricates content cannot meaningfully judge what is safe to input into it, or reliable to take out of it — competence is the precondition for both the confidentiality and the verification duties above, not a separate, softer concern.
A different kind of worry entirely shows up in Manitoba’s and Ontario’s guidance — not accuracy or confidentiality, but who benefits when a tool makes work faster. Manitoba: “fee arrangements should not generate an inappropriate windfall for a lawyer arising from the efficiencies created by using an AI tool.” Ontario, from the billing side specifically: “if a licensee is billing by the hour, they can only charge for the time actually spent by the licensee on the file, even if a generative AI tool has made the task much more efficient.” (LSA, Gen AI Rules of Engagement) This concern has nothing to do with whether AI output is accurate — it is about who owns the value of the time it saves, and both law societies answered the same way: the client does.
Canada’s privacy commissioners frame a version of this worry that applies well beyond any one profession. Developers must evaluate “the training data sets to ensure that they do not replicate, entrench, or amplify historical or present biases” — a concern the principles tie specifically to “highly impactful contexts such as health care, employment, education, policing, immigration, criminal justice, housing or access to finance.” The same principles name explicit “no-go zones”: organizations should not deploy AI systems for “profiling that may lead to unfair, unethical, or discriminatory treatment, or creating outputs that threaten fundamental rights and freedoms.” (OPC, generative AI principles, 2025-05-06)
A three-lawyer firm adopts an AI drafting tool firm-wide without a policy governing its use. Within a month, three separate versions of the same underlying worry surface: an associate pastes an unredacted client agreement into the tool to summarize it (the confidentiality risk); a partner relies on a case citation the tool produced without checking CanLII (the hallucination risk, and in Ontario, now a certification exposure under Rule 06.1); and the firm considers billing every file at the pre-AI hourly estimate regardless of how quickly the tool let it move (the billing-fairness risk both Manitoba and Ontario have addressed directly). None of these requires banning the tool to fix — a short written policy covering redaction, mandatory source verification, and time-based billing addresses all three, which is exactly the kind of response every regulator surveyed here has asked for instead of a prohibition.
Put the five worries together — hallucination, confidentiality, competence, billing fairness, and bias — and a pattern emerges: every one of them is manageable by the person using the tool, given the right discipline. A hallucination is caught by verification against CanLII or a government source. A confidentiality leak is prevented by redacting before input. A competence gap closes with training. A billing windfall is avoided by billing for time actually spent. None of these is a reason the technology cannot be used responsibly — each is a reason it has to be used deliberately. That is the through-line across every Canadian regulator surveyed here, and it is why the guidance keeps expanding while the answer to “is it allowed” keeps coming back the same: yes, with conditions.
Related: can lawyers use generative AI in Canada, regulated professions and AI guidance, and what happens when AI invents a case citation
No. The Law Society of Alberta's survey found that of eleven courts, nine law societies and several other bodies that have issued guidance, none has attempted to ban its use.
Hallucination — fabricated citations, mischaracterized case law, or invented quotations. It appears, under different names, in the Federal Court's, Ontario's and Alberta's court guidance alike.
Because it's a distinct question: even where AI output is completely accurate, Manitoba's and Ontario's guidance holds that the client, not the lawyer, should capture the value of the time an AI tool saves.
The same five risks — verification, confidentiality, competence, fairness and bias — show up in any organization's AI adoption, not only a law practice.
