The Law Society of Ontario has published two documents on generative AI, both in April 2024: a white paper called Licensee Use of Generative Artificial Intelligence and an eight-page practice resource called Generative AI: Your Professional Obligations. Read the practice resource first. Neither document creates new rules. Both apply the Rules of Professional Conduct you already know to a tool that behaves like nothing else in your practice, and the practice resource organizes the analysis into six duties.
The six duties
Competence. You don't have to use AI. If you do, you have to understand it well enough to use it safely, starting with the central failure mode: these tools generate confident, fluent, wrong answers. The LSO's tip is blunt: "Always independently verify any information produced by generative AI that you intend to rely on. The verification process should be completed by a human being, not the AI system itself." Asking the chatbot "are you sure?" is not verification. Competence cuts the other way too: it's the argument for learning these tools properly, not a reason to avoid them.
Confidentiality. The practice resource's worked example: upload a factum into ChatGPT for refinement and the document's content becomes available to the AI engine, where it could be used to train the model. Note what the stated concern is — uncontrolled training and retention, which are exactly the things account settings and commercial terms control. The duty is the same reasonable-steps duty you already apply to every vendor that holds client data, from practice management to document storage: know the tool's terms before client information goes in. The consumer and commercial versions of the same product are different deals. (I've broken down Claude's terms as a worked example.) Where the terms can't protect the information, the LSO's sequence applies: anonymize what you can, and where anonymizing won't adequately protect confidentiality or privilege, explain the risks to your client and get informed consent before using the tool.
Honesty and candour. You don't have to disclose every AI-assisted spellcheck. You do have to inform clients where AI use may affect their interests, the outcome, or the cost of the service. The white paper gives four factors for when to tell the client: whether court rules will force disclosure anyway, whether the client reasonably expects you personally to do the work, reputational risk, and whether you're feeding their information into the tool.
Supervision. The practice resource's most useful sentence: using generative AI tools "is akin to receiving assistance from a non-licensee employee." You can delegate to it. You remain responsible for everything it produces, the same as with an assistant's draft. If you wouldn't file an unread junior memo, don't file unread AI output.
Fees and disbursements. If you bill hourly, you bill for time spent. AI doing a task in minutes means you can't bill the hours it used to take. And if you pass AI tool costs through as a disbursement, bill them at actual cost, not an estimate.
Not misleading the tribunal. Verify citations before they go in a filing, comply with any court-specific AI directions, and keep records of your prompts and verification steps for court-bound work.
The case law
Ontario courts have already dealt with lawyers who skipped verification. In Ko v. Li, a Toronto lawyer faced a contempt proceeding in 2025 after filing a factum with AI-hallucinated case law. By December the matter had produced a referral to the Attorney General. In R. v. Chand, an Ontario court caught fictitious citations, ordered fresh submissions, and barred generative AI from the legal research that followed. In BC, Zhang v. Chen put costs personally on the lawyer who filed two fake ChatGPT cases, and the Federal Court has ordered personal costs for undeclared AI use. The Federal Court now requires a declaration, in the first paragraph of a document, where AI played something like a co-author role. Manitoba and Yukon have their own disclosure directions. Every sanction on the list was for filing unverified output, not for using the tools.
One more date to know: as of January 2026, Legal Aid Ontario's annual self-report requires roster lawyers to confirm they've read and are complying with the LSO's AI guidance. If you do legal aid work, the two documents above are now mandatory reading.
The checklist
What compliance looks like for a small Ontario firm, in ten items:
- Read the eight-page practice resource. It's the regulator telling you what it cares about.
- Inventory what's already in use at your firm, including what staff are using without telling you.
- Check the data terms of each tool. Prefer commercial tiers over consumer tiers for anything touching client work.
- Verify the training and retention settings on every account. Consumer-tier defaults are not in your favour.
- Write a firm AI policy: approved tools, what data can touch them, who verifies what.
- Decide when you'll get informed consent, and update your retainer language where AI use touches client information. Alberta's law society recommends addressing AI in retainer letters outright. The LSO doesn't prescribe wording, but the informed-consent duty points the same direction.
- Build the verification habit: a human checks every citation, factual claim, and number before it leaves the firm.
- Keep records for court-bound work: prompts, sources, what you verified.
- Check the practice directions of every court you appear in. They differ.
- Train yourself and your staff. The competence duty doesn't distinguish between you and the clerk who runs the tool.
BC and Alberta
If you practice in more than one province: BC's law society published its guidance in late 2023 and is more explicit about consent (fully informed and voluntary, in writing or orally with a written record, where confidential information is involved). Alberta's Generative AI Playbook is the most operational of the three. It's the one that recommends firm policies and retainer-letter language outright. The duties rhyme across all three provinces. The paperwork expectations differ at the margins. A cross-Canada comparison is its own post, coming later this summer.
Getting competent
The guidance tells you to be competent with these tools. It doesn't teach you what they do on a real file, where they fail, or what verification looks like in practice. That part is learnable in an afternoon. Inn Laws runs live, hands-on AI training for Canadian lawyers and their staff — vendor-neutral, working on your own files, with the failure modes shown on screen. Sessions count toward CPD in most provinces.