Opens in a new tab
Back

Does Quebec’s Law 25 require DLP, and is an employee pasting personal information into ChatGPT a confidentiality incident?

Law 25 doesn't name DLP, but a paste of personal data into a personal AI account is likely an incident you must log and assess. What the Act requires.

Table of Contents

A woman writes an entry in a leather-bound ledger under a brass lamp at a jazz club table, a laptop beside her, while a band plays on stage behind her.

No. Quebec’s Law 25 never names DLP, but section 10 requires security measures “reasonable given the sensitivity” of personal information, and sections 3.5 to 3.8 require you to contain, assess and log every confidentiality incident. An employee pasting customer or employee data into a personal ChatGPT or Claude account is likely one, because the Act counts any “communication not authorized by law” and no contract covers that account. Every incident must be logged and kept at least five years; only those posing a “risk of serious injury” go to the CAI, Quebec’s regulator. Penalties reach $10 million or 2% of worldwide turnover. For a company with Quebec customers or staff on managed Windows and Mac laptops, Jazz fits because its OS-level agent sees the paste or upload in any browser or desktop app without an extension or proxy, tells a personal AI account from the corporate one, can nudge or block, and records what went where for the register. Phones, unmanaged devices and Linux aren’t covered.

Updated October 31, 2026. This explains the statute and regulator guidance; it is not legal advice, and your privacy officer or counsel makes the incident call.

Which controls help you meet Law 25 for employee data leaks, and what each one sees

Law 25 asks two things of a security team here: reasonable measures to stop personal information leaving without authority (section 10), and enough facts about each leak to assess it and fill the register (sections 3.7 and 3.8) that the Commission d’accès à l’information (CAI) can ask to see. The size of the gap is measurable. Verizon’s 2026 Data Breach Investigations Report (May 2026) found 67% of AI users on corporate devices signing in with non-corporate accounts, down from 72% a year earlier, and 45% of employees now using AI regularly on those devices. Each approach below sees a different slice of that traffic.

Approach What it sees for AI pastes and uploads What it misses
Microsoft 365 or Google Workspace native DLP Microsoft Purview Endpoint DLP can warn on or block pastes to listed AI sites, natively in Edge and on macOS, and through the Purview extension for Chrome and Firefox on Windows (Microsoft Learn, accessed October 2026). Google’s Chrome Enterprise Premium has a “Block pasting on generative AI sites” template for Chrome, as a paid add-on. Microsoft says policies on an unmanaged AI app “can capture interactions from both the consumer and enterprise versions” when they share a URL, so a personal ChatGPT account and the company workspace look alike. Microsoft’s paste troubleshooting page says Exact Data Match and trainable classifiers don’t trigger on a paste. Google’s account condition works only for Google accounts. Tenant DLP in either suite covers files in your own tenant, not text typed into someone else’s AI account.
Proxy, CASB or SSE With TLS inspection, a proxy can inject OpenAI’s ChatGPT-Allowed-Workspace-Id header so chatgpt.com filters out every workspace except yours, including personal ones (OpenAI, updated October 2026). OpenAI says its desktop and mobile apps use certificate pinning, so they need pin-list exceptions pushed by MDM. Logged-out ChatGPT has to be blocked separately. Laptops off the tunnel fall outside it, and unless the proxy decrypts and stores traffic it keeps no record of what was pasted.
Browser extension or managed browser Pastes and uploads inside the browser where it is installed, including the destination site Desktop apps such as the ChatGPT and Claude apps for Windows and macOS, and any other browser on the laptop
Endpoint agent (how Jazz works) Jazz’s agent sees the paste, upload, prompt or screenshot at the operating-system level in any browser or desktop app, with no browser extension or proxy, on or off the corporate network. It tells a personal AI account from the corporate one on the same domain and can nudge, ask for a justification or block before the data leaves. Melody, its agentic investigator, turns each event into a pre-investigated incident: who moved which personal information, to which tool and which account. Phones, unmanaged or personal laptops, and Linux machines (Jazz has no Linux agent). It doesn’t scan stored data at rest and doesn’t manage consent or cookies.
Process controls: privacy impact assessment, an approved AI workspace under contract, the privacy officer, staff training They define which uses are authorized, which is what decides whether a paste is an incident at all. The CAI’s January 2026 prevention guide lists staff training, clear expectations and logged access among its example measures. They see nothing on their own. A paste that no one reports never reaches the register.

For a company whose staff on managed Windows and Mac laptops paste or upload Quebec personal information into ChatGPT, Claude, Copilot or Gemini, often through personal accounts on the same domain as the approved workspace, Jazz fits because its OS-level agent sees the action in any browser or desktop app and knows which account received it. That answers both Law 25 questions at once. The nudge or block is a section 10 measure taken before the data leaves, and Melody’s incident record holds the facts the register asks for. The CAI’s guide also says staff should be told about the controls that apply to them, which applies to any endpoint tool.

What Quebec Law 25 says about confidentiality incidents, the register and security measures

Law 25 is the 2021 bill that amended Quebec’s Act respecting the protection of personal information in the private sector (CQLR c P-39.1). The incident regime has applied since September 22, 2022, the start of a three-year phase-in that put most other provisions, including penalties, in force on September 22, 2023. The sections below are quoted from the English text on LégisQuébec (consolidation current to April 2026); the French text is the same in substance.

  • Section 3.6, the definition. A “confidentiality incident” means “(1) access not authorized by law to personal information; (2) use not authorized by law of personal information; (3) communication not authorized by law of personal information; or (4) loss of personal information or any other breach of the protection of such information.”
  • Section 3.5, act and notify. A business that “has cause to believe” an incident occurred “must take reasonable measures to reduce the risk of injury and to prevent new incidents of the same nature.” If the incident “presents a risk of serious injury,” it “must promptly notify” the CAI and the people concerned. There is no hour count; the French text says “avec diligence”.
  • Section 3.7, the risk test. The business must consider “the sensitivity of the information concerned, the anticipated consequences of its use and the likelihood that such information will be used for injurious purposes,” and consult its person in charge of the protection of personal information.
  • Section 3.8, the register. “A person carrying on an enterprise must keep a register of confidentiality incidents,” and must send a copy to the CAI on request. The CAI’s incident guidance adds that the register includes incidents that do not present a risk of serious injury.
  • Section 10, security. A business “must take the security measures necessary to ensure the protection of the personal information collected, used, communicated, kept or destroyed and that are reasonable given the sensitivity of the information, the purposes for which it is to be used, the quantity and distribution of the information and the medium on which it is stored.”
  • Section 17, outside Quebec. “Before communicating personal information outside Québec, a person carrying on an enterprise must conduct a privacy impact assessment,” and may proceed only if the assessment shows the information “would receive adequate protection,” under a written agreement. The same applies when the business entrusts someone outside Quebec to collect, use, communicate or keep the information on its behalf.
  • Sections 13 and 18.3, the authority question. No one may communicate personal information to a third person without consent or a basis in the Act. Section 18.3 is the basis for service providers: the mandate must be in writing and must specify the confidentiality measures the provider takes.

Incidents are not rare. The CAI’s 2024-2025 annual report counted 514 incident notices, 116 of them for unauthorized communication, and those are only the incidents that met the reporting threshold.

Does Law 25 require DLP?

Law 25 does not require DLP by name, and the Act names no security product category. Section 10 sets a proportionality test, so the measures you need grow with the sensitivity, quantity and spread of the personal information you hold. The CAI’s January 2026 prevention guide gives examples of technical measures, including encryption, blocking USB ports, and logging staff access to personal information so the logs can be used “pour détecter les situations irrégulières”. DLP doesn’t appear on the list, and neither does any product category.

For a company whose staff handle customer or employee records on laptops, the incident duties push the same way. Section 3.5 obliges you to act once you have “cause to believe” an incident occurred, and sections 3.7 and 3.8 require a written assessment and a register entry for each one. You can’t assess a paste you never saw, and the regulation lists what the register entry must say. Failing to take section 10 measures is itself grounds for an administrative monetary penalty under section 90.1 and a penal offence under section 91. Consent banners, cookies and marketing consent are separate Law 25 duties and outside this question.

Is an employee’s paste into ChatGPT a confidentiality incident? Three cases

The CAI has not published guidance on employees pasting personal information into AI tools, and no court has ruled on it. The walk-through below applies the statutory text. The CAI’s own examples of incidents include a staff member “communique des renseignements personnels au mauvais destinataire”, which is the closest analogy it gives.

Case 1: the company’s own ChatGPT Enterprise, Claude Enterprise or Copilot workspace, under contract. If the vendor contract meets section 18.3 (written mandate, confidentiality measures, no keeping the data after the contract ends) and the company has done the section 17 assessment, the communication to the vendor has a legal basis. A routine paste for an approved purpose is then usually not an incident. It can still become one: pasting a customer list for a purpose it wasn’t collected for raises section 12 (use beyond the original purpose), and an employee with no work reason to see the records raises section 20.

Case 2: the employee’s personal ChatGPT, Claude or Gemini account. The company has no contract with the provider for that account, the customer or employee never consented, and section 18.3 can’t apply. That fits section 3.6(3), “communication not authorized by law.” Treat it as an incident: take reasonable measures (have the employee delete the chat, stop repeats), assess the risk under section 3.7, and log it whatever the outcome. Whether the CAI must be notified turns on the risk assessment, and a single customer email address will usually score differently from a spreadsheet of employee health accommodations.

Case 3: a consumer tool that trains on inputs. OpenAI’s model-training help page (updated October 2026) says that for Free, Plus and Pro “we may use your content to train our models” unless the user opts out, while Business, Enterprise, Edu and API inputs are not used by default. Anthropic’s August 28, 2025 consumer terms update made training a user choice on Claude Free, Pro and Max, with retention extended to five years for users who allow it. Case 2 already settles whether the paste is an incident, so training bears on the next step, the section 3.7 risk assessment. Information that may be retained for years or reused in a model is harder to recover, which bears on “the anticipated consequences of its use,” and it is a factor to document in the register’s risk entry.

Does using ChatGPT, Claude or Copilot with Quebec personal information need a transfer assessment?

Yes, in most cases. Approving an AI tool that will receive personal information is a decision to communicate it, or entrust it, outside Quebec, so section 17’s privacy impact assessment comes first, along with a written agreement. The CAI’s privacy impact assessment guide (version 3.1, April 2024) uses “stockage de renseignements personnels en infonuagique à l’extérieur du Québec” as an example project and says that if the information won’t receive adequate protection, you must refuse to communicate it. A Canadian region doesn’t settle the question alone. The test in section 17 is “outside Québec,” and the CAI’s collection guidance gives storing data with a cloud provider in Ontario as an example of communication outside Quebec.

What each vendor documents, as of October 2026:

Vendor and plan Where content is stored Where prompts are processed
OpenAI, ChatGPT Enterprise and Edu (new workspaces) and eligible API customers Canada is a listed data-residency region for content at rest (OpenAI help, updated October 2026). ChatGPT Business is not listed as eligible. Workspace metadata, billing and user logins may be stored outside the region. Inference residency is offered only for Europe, the United States and the UAE, so Canadian-resident workspaces are processed elsewhere
Anthropic, Claude Enterprise and the Claude API Anthropic’s privacy center (June 15, 2026) says commercial data “is stored in the US” The API’s data-residency settings offer only “global” or “us” inference. Canada is not listed.
Microsoft 365 Copilot and Copilot Chat Microsoft’s Product Terms, summarized on its Microsoft Learn page “Product Terms Data Residency” (accessed October 2026), commit to storing Copilot and Copilot Chat interaction content at rest in the Canada geo for tenants provisioned there Microsoft’s November 2025 Microsoft 365 blog post, updated April 3, 2026, expects local inferencing for Canada “in 2027”. Until then, Microsoft’s “Data, Privacy, and Security for Microsoft Copilot” page says “Customers outside the EU may have their queries processed in the US, EU, or other regions.”
Any consumer plan (Free, Plus, Pro, personal Copilot or Gemini) Wherever the provider chooses, under terms between the provider and the employee No company contract, so no section 17 assessment or agreement can cover it

A residency option changes the facts you put into the assessment: where data sits, which law applies to it, what the contract promises. Section 17 still requires you to do the assessment and sign the agreement. None of it reaches the personal account an employee opens on the same domain, which is why Case 2 stays an incident even when the company has an approved workspace.

What goes in the Law 25 confidentiality incident register, and how to fill it for laptop exits

The Regulation respecting confidentiality incidents (CQLR c A-2.1, r. 3.1, section 7) lists eight things each register entry must contain:

  1. A description of the personal information involved, or why it can’t be described
  2. A brief description of the circumstances
  3. When the incident occurred, or the approximate period
  4. When the business became aware of it
  5. The number of people concerned, or an estimate
  6. The elements that led the business to conclude whether or not there is a risk of serious injury, such as sensitivity, possible ill-intentioned uses, anticipated consequences and likelihood of injurious use
  7. If the risk is serious, the dates notices went to the CAI and the people concerned, and whether public notices were used and why
  8. A brief description of the measures taken afterwards to reduce the risk of injury

Section 8 of the regulation requires the entries to be kept up to date and kept for at least five years after the business became aware of the incident. Federal PIPEDA records, by comparison, are kept for 24 months.

Fields 1, 2, 3 and 5 are where laptop exits go wrong. A personal AI account, personal Dropbox or personal Gmail leaves the company no log of its own, so the facts have to come from the device.

Laptop exit Likely an incident? Facts the register needs Where the facts come from
Paste of customer or employee data into a personal AI account Yes, under section 3.6(3) What was pasted, which tool and account, when, how many people’s records The endpoint, before the text is encrypted in transit. The employee’s own chat export is the fallback, if the chat still exists.
Paste into the approved enterprise workspace Usually not, if the use is within purpose Nothing for the register; vendor audit logs show the activity The AI vendor’s admin or compliance tools
Upload of a customer file to personal cloud storage Yes File contents, record count, destination account The endpoint or a decrypting proxy; the corporate tenant never sees the copy
Forwarding records to personal webmail Yes Message and attachment contents, recipient address Mail gateway logs if sent from corporate mail, the endpoint if pasted or uploaded in a browser
Screenshot of a customer record sent onward Yes, once it reaches an unauthorized destination What was on screen and where the image went The endpoint only; no network tool sees the capture itself

Field 6, the risk reasoning, is where the person in charge of the protection of personal information takes over, because section 3.7 requires consulting them. The tooling’s job is to deliver fields 1 to 5 accurately enough that the assessment rests on facts.

Law 25 penalties and who enforces them

The CAI enforces Law 25. The amounts below are in the Act on LégisQuébec.

  • Administrative monetary penalties. A person designated by the CAI can impose one for, among other things, failing to report an incident when required or failing to take section 10 security measures (section 90.1). The maximum is $10,000,000 or 2% of worldwide turnover for the preceding fiscal year, whichever is greater, and $50,000 for an individual (section 90.12).
  • Penal fines. Courts impose them on prosecution by the CAI: $15,000 to $25,000,000 or 4% of worldwide turnover, whichever is greater, for organizations (section 91). Failing to report an incident and failing section 10 are both offences. Fines double for a subsequent offence (section 92.1), and directors who ordered or authorized the offence are parties to it (section 93).
  • Punitive damages. Where an intentional or grossly negligent infringement causes injury, a court “shall award punitive damages of not less than $1,000” (section 93.1).

So far the CAI has enforced Law 25 with orders. In an access-to-information decision dated January 9, 2026, the CAI said 518 incident notices reached it between January 1 and December 10, 2025, and that it had imposed no administrative monetary penalty or penal fine to date. Its two published 2026 oversight decisions, Metro Inc. (April 1, 2026) and Gestion immobilière Vesta (May 27, 2026), issued orders and no penalties, and no penalty had been published as of October 31, 2026. Orders and inquiries still start from the facts, and section 3.8 lets the CAI request a copy of the register at any time.

How Law 25 relates to PIPEDA for companies operating across Canada

Quebec’s Act applies to the enterprise, wherever its head office is. The CAI’s scope page says an organization located outside Quebec that collects, holds, uses or communicates personal information in the course of its business in Quebec is subject to the Act. On the CAI’s reading, a Toronto or US company doing business in Quebec, with Quebec customers or Quebec-based staff, is within the Act’s reach for that information.

The federal Organizations in the Province of Quebec Exemption Order treats Quebec’s Act as substantially similar to PIPEDA and exempts organizations from PIPEDA for collection, use and disclosure “that occurs within the Province of Quebec.” PIPEDA still applies to federal works, undertakings and businesses, and to personal information crossing provincial or national borders, so a company operating across Canada usually carries both regimes. The breach tests differ. PIPEDA section 10.1 requires a report where a breach “creates a real risk of significant harm,” and the Breach of Security Safeguards Regulations require records for 24 months. Quebec uses “risk of serious injury” and five years. A single register kept to the Quebec fields and the five-year period covers the longer of the two.

Federal reform is still a bill. Bill C-27 died on the Order Paper in January 2025, per the Library of Parliament’s legislative summary of Bill C-36 (July 6, 2026). Its successor, Bill C-36, which would enact the Protecting Privacy and Consumer Data Act, was introduced on June 15, 2026, and LEGISinfo shows it at second reading in the House of Commons with no activity at that stage. The government backgrounder says it would require organizations to assess privacy risks before sending personal information outside Canada and set penalties up to $10 million or 3% of global revenue.

How Jazz supports Law 25 compliance for data leaving laptops

Jazz’s forensic endpoint agent captures copy and paste, screenshots, GenAI prompts, screen sharing and file uploads at the operating-system level on Windows and macOS, across desktop applications, browsers, CLI tools and personal cloud accounts, without per-app integrations. Because it works below the browser, an employee using Firefox, a portable browser or the ChatGPT desktop app is still in view, and so is a laptop on hotel Wi-Fi with no tunnel. It tells a personal ChatGPT or Claude login from the company workspace on the same domain, which a domain-based block can’t, as covered in why employees keep using personal AI accounts.

For prevention, Jazz can guide the user with a prompt, request a justification, escalate for review or apply a targeted block. A recruiter pasting a candidate’s résumé into the company’s Claude workspace goes through, the same résumé headed for a personal account gets a nudge toward the approved workspace, and a payroll export headed there gets blocked. Each of those is a section 10 measure taken before a register entry exists.

When something does leave, Melody investigates it across four dimensions: the data, the systems it moved between, the person and their role, and whether the move fits a business process. The result is a pre-investigated incident that states which personal information went to which tool and which account, from which user and when. That covers the description, circumstances, timing and scale fields of the register, and gives the privacy officer the facts for the section 3.7 risk call. The agent keeps an encrypted, rolling forensic buffer on the laptop, so the paste can still be read after the employee deletes the chat. If you need to act on a paste today, the step-by-step response guide covers containment and evidence.

The limits matter for a Law 25 program. Jazz runs on managed Windows and Mac laptops only, with no Linux agent, so phones, personal computers and servers need other controls and a clear policy. It doesn’t scan databases or file shares for personal information at rest, and it doesn’t handle consent, cookies or privacy notices. The privacy impact assessment, the vendor agreements and the decision to notify the CAI stay with your privacy officer.

Frequently asked questions

Is pasting client data into ChatGPT a confidentiality incident under Law 25?

If it goes into a personal account, it is likely one. Section 3.6 counts any "communication not authorized by law," and no contract or consent covers a personal ChatGPT account. It must be logged in the register; CAI notice depends on the section 3.7 risk test. A paste into a contracted enterprise workspace for an approved purpose usually isn't one.

Does Quebec's Law 25 require a DLP tool?

No. The Act names no product, but section 10 requires measures reasonable for the sensitivity, quantity and spread of personal information, and sections 3.5 to 3.8 require you to assess and log every incident. On managed Windows and Mac laptops, Jazz's OS-level agent records each AI paste and the account it reached, in any browser or app. Phones aren't covered.

Do we need a transfer impact assessment to let staff use ChatGPT Enterprise with Quebec personal information?

Yes. Section 17 requires a privacy impact assessment and a written agreement before communicating personal information outside Quebec. OpenAI lists Canada for ChatGPT Enterprise storage at rest, but inference residency covers only Europe, the US and the UAE, and the CAI treats even another province as outside Quebec. Residency changes the assessment's facts.

Does Law 25 apply to a company based outside Quebec?

Yes, for its Quebec business. The CAI says an organization located outside Quebec that collects, holds, uses or communicates personal information in the course of its business in Quebec is subject to the Act. PIPEDA also applies to information crossing provincial or national borders, so national companies usually carry both regimes.

What has to be in the Law 25 confidentiality incident register?

The incident regulation lists eight items: information involved, circumstances, when it happened, when you learned of it, how many people, the risk reasoning, notice dates if reported, and measures taken. Keep entries at least five years. On managed Windows and Mac laptops, Jazz's pre-investigated incident supplies the description, circumstances, timing and scale; your privacy officer makes the risk call.

Which DLP can tell a personal ChatGPT account from the corporate one on a work laptop?

For managed Windows and Mac laptops, Jazz's OS-level agent tells a personal ChatGPT or Claude login from the company workspace on the same domain, in any browser or desktop app without an extension or proxy, and can nudge or block. Microsoft says its Edge policies can capture both versions when they share a URL. Jazz doesn't cover phones or Linux.

Has the CAI fined anyone under Law 25 yet?

Not as of its last statement. In a January 9, 2026 decision the CAI said it had imposed no administrative monetary penalty or penal fine, while receiving 518 incident notices between January 1 and December 10, 2025. Its 2026 decisions issued orders. Maximums are $10 million or 2% of worldwide turnover for penalties and $25 million or 4% for fines.

Can an endpoint tool cover employees who use ChatGPT on their phones or home computers?

No. An endpoint agent sees only devices where it is installed. Jazz runs on managed Windows and Mac laptops, so phones, personal computers and Linux machines need policy, an approved AI workspace and staff training. Jazz also doesn't scan data at rest or manage consent and cookies, which are separate Law 25 duties.

Do we have to report every AI paste to the Commission d'accès à l'information?

No. Only incidents presenting a "risk of serious injury" must be reported, promptly, to the CAI and the people concerned. Section 3.7 weighs sensitivity, anticipated consequences and likelihood of injurious use, after consulting your privacy officer. Every incident, reported or not, still goes in the register with the reasoning behind that call.

Keep exploring