The situation
At a 30-lawyer Montreal firm, a lawyer has to send a demand letter before the end of the day. She pastes her draft into ChatGPT to tighten the wording. The text contains the client's name, the opposing party's name, the amount claimed, and the facts at issue. At the same time, a legal technician is having a shareholders' agreement translated into English for an Ontario client. Nobody feels they are communicating anything.
The managing partner, who has just registered for the Barreau's new mandatory course on generative AI, is asking the question the whole firm should ask: what do the Barreau and professional secrecy actually allow, and how would we prove it if a client asked?
What the rules actually say
The practical guide and mandatory training
On October 25, 2024, the Barreau du Québec published its practical guide to the responsible use of generative AI (in French), since reissued in a second edition. The Barreau encourages generative AI use, but recommends it only where the risks can be managed effectively. A FAQ and a toolkit complement it. Since April 1, 2026, a two-hour mandatory course must be completed by every member before April 1, 2027.
The guide and FAQ are guidance. The binding obligations are in legislation and the Code of Professional Conduct, which the guide ties to each best practice.
Professional secrecy and confidentiality
Section 9 of the Charter of human rights and freedoms guarantees every person the right to professional secrecy, and section 60.4 of the Professional Code requires every professional to preserve it. The Code of Professional Conduct of Lawyers adds, in section 60, that a lawyer must ensure the confidentiality of all information about a client's affairs and, in section 61, that the lawyer must take reasonable measures so that every collaborator, and the firm, protects that information.
The guide takes a clear position. It states that simply entering information protected by professional secrecy into an open system, meaning one that is publicly accessible and does not keep data in-house, is a breach, even without any actual reproduction or disclosure. It recommends anonymizing before any input, including indirect details such as job titles or specific places, limiting data to what is strictly necessary and, if anonymizing is impossible, not using the tool.
Competence and supervision
Since 2020, section 21 of the Code has included knowledge and skills related to the information technologies a lawyer uses. The guide concludes that lawyers must understand each tool, its limits and its terms before relying on it. Section 35 makes the lawyer responsible for the mandate and requires adequate supervision of anyone collaborating on it, and section 5 requires reasonable measures so that collaborators respect the rules. According to the FAQ, that supervision extends to how the lawyer and the team use AI tools.
Law 25
A law firm is also an enterprise under the private sector act. Section 10 requires reasonable security measures, and section 17 requires a privacy impact assessment before personal information is communicated outside Quebec, followed by a written agreement. A prompt containing a client's name, sent to a tool operated outside Quebec, falls within that framework. Our post on Law 25 and AI prompts and our page on Quebec Law 25 cover the regime in more detail.
The courts
On October 24, 2023, the Superior Court issued a notice to the legal community and the public on the integrity of submissions when large language models are used, urging caution and rigorous human verification of cited sources. The Court of Québec and the Court of Appeal have also issued notices on AI use.
Why policies and bans fall short
A policy that says "anonymize before you paste" rests on a fragile assumption: that every person, under pressure, will spot every direct and indirect identifier in a three-page text. The guide itself notes that a job title or a location can be enough to identify someone.
Banning the tool does not solve the problem either. Use shifts to phones and personal accounts, outside any supervision. LayerX found in 2025 that 71% of generative AI connections go through personal rather than corporate accounts. And on the Barreau's reading, the breach is complete at the moment of input, so there is nothing to fix afterward. Once submitted, content cannot be recalled and becomes subject to the provider's terms, which can permit retention, sub-processing and, in some cases, training.
What a practical control looks like
The Barreau's guide supplies most of the elements of a program:
- Adopt an internal generative AI policy. Approved tools, mandatory firm accounts, and data that must never be entered: client names and identifiers, facts of the file, deal terms, personal information.
- Assess each tool. Read the terms of use, check data retention and reuse, and complete the section 17 assessment where personal information may be communicated outside Quebec.
- Anonymize and minimize, or abstain. If anonymizing is impossible, the public tool is not an option.
- Structure client communication. Explain the use of AI, its benefits and risks, and record informed consent where the situation calls for it.
- Train the whole team. The mandatory course covers lawyers; legal technicians, assistants and articling students need concrete guidance too.
- Plan the incident response. The guide and FAQ recommend acting quickly, informing clients and documenting everything; also assess your Law 25 obligations with your privacy officer.
- Verify sources separately. References cited to a court are checked against reliable sources, by a lawyer.
For what to do when a file has already been pasted into a public tool, see our guide to the next 48 hours.
Sanitized Ai is a browser extension that acts on steps 1, 3 and 5 at the exact moment the risk arises. When someone pastes or uploads client names and identifiers, personal information, or deal terms into an AI assistant, it detects that data and redacts or blocks it before submission. It then explains to the person, in plain language, what was flagged and why. Coverage spans the major AI assistants and keeps growing.
Administrators get a dashboard of flagged events (which tool, what type of data, which policy, when) that never shows prompt content. That audit-ready record can help demonstrate the reasonable measures section 61 requires, without anyone at the firm reading information protected by professional secrecy. It does not check citations or legal accuracy. Confirm your approach with the Barreau or your own counsel, and see how the extension fits a firm's workflow on our law firms page.