Disclosure: DROZlegal publishes this article and builds a practice-automation product for Canadian law firms. The privilege analysis below draws on published commentary from Thomson Rogers LLP, Norton Rose Fulbright, and Cain Lamarre — named and dated throughout — cross-checked against each firm's own published text this week. This is general legal information, not legal advice; verify current guidance directly with those sources or your own counsel before relying on it.
Why this question doesn't have a settled answer yet
Privilege law was written for people, and AI tools are not people. Solicitor-client privilege protects communications that meet three conditions: the communication is with a lawyer, it's made for the purpose of seeking or giving legal advice, and it's intended to stay confidential. An AI chatbot fails the first test on its face — it isn't a lawyer — and most consumer AI terms of service don't promise the confidentiality the third test requires.
Thomson Rogers LLP partner Robert M. Ben put it directly in a March 3, 2026 firm article: client-initiated AI chats on consumer tools "are not private, confidential, or subject to any form of legal privilege," and could be producible to an opposing party in a lawsuit if relevant. His practical advice to clients is blunt — never paste a lawyer's advice, strategy, or settlement position into a consumer AI tool, because that is treated as voluntary disclosure of privileged material to a third party and likely amounts to waiver.
Litigation privilege is a different, broader test
Litigation privilege is the one more likely to actually cover AI-assisted research. It protects material prepared for the "dominant purpose" of existing or reasonably anticipated litigation, and Canadian courts have built it into a doctrine that's broader than the US work-product protection it's often compared to — it can extend to non-lawyers, self-represented litigants, and material that isn't itself confidential, so long as the dominant-purpose test is met.
The doctrine's foundations come from the Supreme Court of Canada, not from anything AI-specific: Lizotte v. Aviva Insurance Company of Canada, 2016 SCC 52, and Blank v. Canada (Department of Justice), 2006 SCC 39, are the cases Norton Rose Fulbright cites as the general dominant-purpose authority in its June 2026 analysis of privilege and generative AI. Neither case involves AI at all — they're cited for the underlying test, not as AI precedent, and no Canadian decision has yet applied that test to an AI tool.
"Privilege depends on confidentiality, control and the structure of the solicitor-client relationship." — Marc-Alexandre Poirier and Badr Belhad, Cain Lamarre, "Solicitor-Client Privilege in the Age of Public AI (Artificial Intelligence) Tools," July 20, 2026
That's the line worth remembering: litigation privilege can potentially reach AI-assisted research and drafting, but only if the confidentiality and control that privilege has always required are actually maintained around it.
The waiver risk: what "disclosure to a third party" means when the third party is an AI vendor
Waiver doesn't require intent — it usually just requires disclosure outside the privileged relationship. All three sources reviewed for this article converge on the same distinction: a public, consumer-facing AI tool carries materially more waiver risk than a tool used under a lawyer's direction inside a confidentiality-controlled environment.
| Factor | Public/consumer AI tool | Firm-controlled/enterprise AI tool |
|---|---|---|
| Who directs the use | Client or staff acting independently, often without the lawyer's knowledge | Used under counsel's supervision, for a legal-advice or litigation purpose |
| Confidentiality terms | General consumer terms of service; often silent or permissive on retention/training | Business contract with stated confidentiality, retention, and training terms |
| Waiver risk on matter-specific facts | Higher — treated by commentators as a serious waiver risk | Lower, but not zero — still fact-specific and untested by any Canadian court |
Norton Rose Fulbright's own recommendation reflects that gap: "As a best practice, litigants should minimize disclosing prejudicial or confidential information to publicly available Gen AI platforms to minimize any risk of inadvertent waiver of privilege."
What Canadian courts and commentators have said so far
The honest answer is that this is unsettled law, not settled law with a surprising result. Thomson Rogers (March 2026), Norton Rose Fulbright (June 2026), and Cain Lamarre (July 2026) each independently reach the same conclusion in their own words: no Canadian court has directly ruled on whether using a public AI tool waives solicitor-client or litigation privilege.
What's driving the commentary right now is a pair of US federal decisions from February 2026 — Heppner and Warner — where courts confronted privilege and work-product claims over AI-generated material. Canadian commentators are watching them closely, but they're US authority applying US work-product doctrine, which is narrower than Canadian litigation privilege in some respects. Treat them as an early signal of how courts are starting to think about the problem, not as a preview of how a Canadian court will rule.
One more layer worth knowing about, even for an Ontario-focused practice: Cain Lamarre notes that Quebec's Charter of Human Rights and Freedoms, s.9, provides a separate statutory professional-secrecy protection that doesn't have a direct Ontario equivalent — a reminder that "Canadian law" on this question isn't one uniform rule across provinces.
This evidentiary question sits next to, but is distinct from, the Law Society of Ontario's confidentiality rule (Rule 3.3-1), which governs a lawyer's professional obligation to understand how an AI tool handles client data before using it — regardless of whether privilege is ever tested in court. We've covered that rule, and the LSO's other generative-AI guidance, in full in our explainer on the LSO's generative AI guidance.
A practical checklist for AI-assisted legal research today
Until a Canadian court weighs in, the commentary above points to the same handful of practical habits:
- Never paste a lawyer's own advice, strategy, or settlement position into a public AI tool. That's the single clearest waiver trigger every source flags.
- Keep AI-assisted research and drafting under a lawyer's direction, tied to a legal-advice or litigation purpose — not a client or staff member experimenting independently on the file.
- Do it inside a confidentiality-controlled environment — a firm-tenant or enterprise tool with a business contract, not a consumer chatbot's default settings.
- If a client wants to organize their own thoughts with AI help, have them do it offline — a private document sent directly to counsel, not a chat with a third-party AI platform.
- Know a tool's training and retention terms before any matter-specific fact goes in. This is a confidentiality-control factor a court would likely weigh, not a guarantee either way.
Where a vendor's data-handling terms fit into this analysis
A vendor's contract terms are one input into a privilege analysis, not a substitute for one. The same four questions worth asking any AI vendor about training use, retention, processing location, and unsupervised authority — covered in full in our vendor-evaluation checklist — are also the confidentiality-control facts a court would likely look at if a waiver argument ever reached a Canadian courtroom on this question.
DROZlegal's own posture, stated plainly: stored client data never leaves Canada (AWS ca-central-1); AI processing runs under Anthropic's commercial terms, meaning no training on client data and auto-deletion within roughly 30 days. That's a factor a lawyer evaluating AI tools may want to weigh in their own confidentiality-control analysis — it is not a guarantee that any specific privilege claim would hold. Privilege is always fact-specific, decided by a court on the facts of a particular file, not established in advance by any vendor's marketing copy, ours included.
Confidentiality and waiver are one axis of AI research risk; accuracy is a separate one. If a research tool's output ends up in a filing, whether it answers from a verified case-law corpus or an open-ended model's recall is its own question — see our guide to AI legal research software with verified case-law citations for what that distinction requires in practice.
The confidentiality-control question above applies just as directly to the major commercial research platforms Canadian lawyers already use. CoCounsel, Thomson Reuters' AI layer on Westlaw and Practical Law, is genuinely sold and used in Canada today — but as of this writing, Thomson Reuters' own Canadian product materials don't state where a Canadian firm's CoCounsel data is actually hosted, which is exactly the kind of processing-location fact worth confirming in writing before a matter-specific fact goes in. We cover that gap, and what it does and doesn't tell you, in our CoCounsel alternative comparison for Canadian law firms.
Not ready to subscribe? Join the DROZlegal waitlist instead.