Canada: Rules for Licensing Data for AI Training

Canadian law on AI training data has no statute written for training: the Copyright Act, which has no text and data mining exception, and PIPEDA's consent rules decide most deals, with Quebec adding a provincial layer and Bill C-36 pending. A full-text search of the Copyright Act on 2026-11-09 found no such exception, and PIPEDA section 7.2(4) withdraws its business-transaction exception where the main purpose of a deal is selling or leasing personal information.

This page sets out the instruments, dated changes and seller-side programs as of 2026-11-09, for the data holder and for the AI developer. It is a free-to-read part of the wider reference on AI training data licensing, and every rule cites its official text. Federal and provincial regulators published joint findings on OpenAI's training data on 2026-05-06, and Bill C-36 had its first reading on 2026-06-15.

Each figure is dated and checked against the official text, as described in the method for sourcing and dating every figure.

What Canadian Law Decides Before a Company Licenses Data for Training

Canadian law decides four things before a data license: whether protected works in the set may be used without permission, whether personal information may be disclosed without consent, whether a provincial law adds duties, and how cross-border payments are taxed. The four gates map to four instruments.

• Copyright: the Copyright Act, R.S.C. 1985, c. C-42, section 29.
• Personal information: PIPEDA, S.C. 2000, c. 5, section 7.2(4).
• Provincial layer: Quebec's Law 25, S.Q. 2021, c. 25.
• Tax: Income Tax Act, R.S.C. 1985, c. 1 (5th Supp.), section 212(1)(d).

The pending federal privacy bill is Bill C-36, not Bill C-27. Rules on employee data and recording consent are not recorded on this page.

Canadian Instruments That Apply to a Training Data Deal

Nine Canadian instruments, findings and record gaps shape a training-data license, and the table gives each citation, its date and whether it is final, pending or open. Rows with a page of their own link to it, and "open" rows mean no primary source was verified on 2026-11-09, not that no rule exists.

How Each Canadian Rule Reaches the Seller and the Buyer

The table lists each instrument; the sections below show what each one asks of the data holder and of the AI developer.

No Text and Data Mining Exception: Fair Dealing and the License Route

The Copyright Act (R.S.C. 1985, c. C-42) contains no text and data mining or informational-analysis exception, so training on protected works needs either a license or a fair-dealing defense. Section 29 provides that fair dealing for the purpose of research, private study, education, parody or satire does not infringe copyright, and a full-text search on 2026-11-09 returned no hits for "data mining", "informational analysis" or "computational analysis". This is the statute's text, not a court holding on training.

For a seller, content it owns or controls has value because no statutory exception gives it away, whereas the EU offers an opt-out regime and Japan Article 30-4 a broad exception with a proviso. The first check is whether the holder owns the works at all, which a rights audit settles before licensing.

Personal Information in the Set: Consent Under PIPEDA

PIPEDA (S.C. 2000, c. 5) makes consent the basis for disclosing personal information in a commercial deal, so a dataset with customer or employee details needs consent or a statutory exception before it is licensed. The Act is in force and the Office of the Privacy Commissioner of Canada (OPC) oversees it, while Bill C-36 would replace its Part 1.

• Do customer records identify individuals, and was consent given for disclosure to a third party?
• Do communications such as emails or chats carry names or contact details of people outside the company?
• Do recordings capture voices of individuals who agreed only to a service call?

De-identification before licensing reduces what PIPEDA reaches, and the pending bill would add an anonymization definition and an exception for research on de-identified data. The practical steps are set out under PII removal and de-identification before a license.

Why the Business-Transaction Exception Rarely Covers a Data License

PIPEDA counts licensing an organization's assets as a business transaction, but it withdraws the consent exception where the main purpose or result of the deal is selling or leasing personal information, which describes most training-data licenses. Bill C-36 would add a business-transaction exception that requires de-identification, and it is pending.

• Section 2(1) defines business transaction to include "the lease or licensing of any of an organization's assets".
• Section 7.2(4) provides that the exception does not apply to a transaction whose primary purpose or result is the purchase, sale or other acquisition or disposition, or lease, of personal information.

Regulator Findings on Scraped and Licensed Data (2026)

In PIPEDA Findings #2026-002, published 2026-05-06, the federal commissioner and the Quebec, British Columbia and Alberta offices found that OpenAI had no implied consent for web-scraped or licensed data and that licensed datasets require that individuals were given a choice. The training-data complaint was well-founded and conditionally resolved, and indiscriminate scraping does not meet the publicly-available exception. OpenAI stated that licensed data were less than 1% of its pre-training datasets, a claim the offices could not verify.

Provinces With Their Own Private-Sector Rules, Starting With Quebec

Quebec's Law 25 (S.Q. 2021, c. 25, Bill 64, assented 2021-09-22) adds consent for each specific purpose and a privacy impact assessment before personal information leaves Quebec. As of 2026-11-09 the Quebec regulator lists penalties up to CAD 10,000,000 or 2% of worldwide turnover (administrative) and CAD 25,000,000 or 4% (penal), whichever is greater. Other provincial private-sector acts are not recorded on this page, and the Quebec page holds the depth.

Cross-Border Payments to and From Canadian Data Holders

A Canadian buyer paying a royalty to a non-resident seller withholds 25% Part XIII tax under Income Tax Act section 212(1)(d) unless a treaty reduces it. The IRS treaty table (Table 1, revised May 2023) lists 0% US withholding on copyright and industrial royalties paid to Canadian residents (income codes 12 and 10), subject to the treaty's limitation-on-benefits test. Sales-tax treatment of licenses to non-residents is not stated here, and the withholding page compares rates.

Enacted, Pending and Proposed: Canadian Changes as of November 2026

Six dated items changed or may change Canadian rules for training data, and the table keeps what is in force apart from a bill at second reading.

Bill C-36 is a bill, not law: its penalties apply only if it is enacted. Section 114 sets a maximum penalty of the greater of CAD 10,000,000 and 3% of gross global revenue, and section 145 sets indictable fines of the greater of CAD 25,000,000 and 5%. Findings bind the parties, not the market, and court rulings on training data are tracked in the AI copyright lawsuits over training data tracker.

Selling Company Data to AI Companies From Canada: Programs That Name Canada

Seven programs and intermediaries name Canada on their own pages, each as a claim checked 2026-11-09, and none of them changes what Canadian law requires of the seller.

Two contradictions stay visible: Mode lists Canada on its referral page and agreement but calls the US the strongest fit on its home page, and Nyne asks for 20+ U.S. W-2 employees but offers Canada in its estimator. Two Canadian companies, Cohere and TELUS Digital, have no public company-data program on record as of 2026-11-09. Endless and McCarthy's AI handle licensing for sellers and do not buy data. Terms across all of them are compared in the company data partnership programs page.

Benefits, Risks, Downsides and Alternatives of Licensing Data Under Canadian Law, Including Not Selling

Canadian law lets a company license data it holds the rights to, and the missing TDM exception gives that content license value, but PIPEDA consent and the provincial layer add conditions, and not licensing remains a valid choice.

Benefits

• No statutory exception gives protected works away, so owned content carries license value.
• A license documents the individual choice that the 2026 findings require.
• De-identified or non-personal data falls outside most consent questions.

Risks

• A license of personal information without consent fails the section 7.2(4) test.
• Regulator findings such as #2026-002 can reach scraped and licensed data alike.
• Quebec penal fines reach CAD 25,000,000 or 4% of worldwide turnover, whichever is greater, as of 2026-11-09.

Downsides

• Consent collection, impact assessments and written agreements cost time.
• Part XIII withholding reduces cross-border income unless a treaty applies.

Alternatives

• License only content that carries no personal information.
• License de-identified data.
• Wait for Bill C-36, which is at second reading.
• Not selling or licensing at all, among the alternatives to selling data to AI labs.

Beyond the Canadian Layer: Other Jurisdictions, Topic Comparisons and the Sale

Canadian law stops at the border and at the contract; the pages below compare topics across jurisdictions, cover the neighboring US layer and apply the rules to a sale.

• AI training data law by jurisdiction, the parent overview.
• United States AI training data law, the federal and state layer next door.
• copyright and AI training across jurisdictions, how copyright exceptions compare.
• employee data and AI training rules by jurisdiction.
• call and meeting recording consent rules compared.
• is web scraping for AI training legal, scraping law across jurisdictions.
• sell or license company data to AI companies, the steps of a sale.
• privacy law and AI training data compared across jurisdictions.
• United Kingdom rules for licensing data, another common-law jurisdiction.

How This Page Is Sourced and Dated

Every rule here cites the consolidated federal statutes, the LEGISinfo record, a regulator's findings or the Quebec regulator's copy of the act, with the date it was checked. The rules for dating and checking are set out in the method for sourcing and dating every figure. This page is general information about Canadian law, not legal advice for any specific deal.

Frequently Asked Questions

Does Canada have a text and data mining exception for AI training?

No, the Copyright Act has no text and data mining exception, only section 29 fair dealing for listed purposes. The full text was searched on 2026-11-09.

Is consent needed to license personal information for AI training in Canada?

Yes, unless an exception applies. PIPEDA section 7.2(4) removes the business-transaction route where a deal's main purpose or result is selling or leasing personal information.

Which federal privacy bill is pending in Canada?

Bill C-36, the Protecting Privacy and Consumer Data Act. It had first reading on 2026-06-15 and was at second reading as of 2026-11-09.

Does publicly available personal information escape consent in Canada?

Not when it is scraped indiscriminately. PIPEDA Findings #2026-002 found that such scraping does not meet the publicly-available exception.