AI Training Data and Copyright Risk
AI training data raises unresolved copyright questions in Canada. Copying works to train models is infringement unless an exception applies, Canada has no text and data mining exception, and fair dealing is untested for commercial training. Licensing and dataset governance are the practical safeguards.

Key takeaways
- Reproducing a substantial part of a work for training infringes copyright unless an exception applies.
- Canada has no specific text and data mining exception in its Copyright Act.
- Fair dealing for research is broad but untested for commercial model training.
- The temporary reproduction exception is unlikely to cover retained datasets.
- Dataset registers, licences and provider indemnities are key risk controls.
AI training data and copyright: the short answer
AI training data is the most contested copyright issue in technology today, and Canadian law has not yet answered it. Training a machine learning model usually involves copying works, such as text, images, code or music, into datasets and processing them. Under s. 27(1) of the Copyright Act, reproducing a substantial part of a work without consent infringes copyright unless an exception applies. Canada has no specific text and data mining exception.
Developers therefore rely on general defences, chiefly fair dealing for research or private study under s. 29, and the exception for temporary reproductions in technological processes under s. 30.71, both of which are untested for large-scale commercial training. Licensing, dataset governance and contractual protections are the practical answers while the law develops.

For AI developers, enterprises fine-tuning models, and businesses buying AI tools, the exposure is real: infringement claims, injunctions, statutory damages for registered and unregistered works, and contractual claims from data providers. As at the Copyright Act consolidation current to 19 October 2026, there were no AI-specific amendments. This article explains how copyright applies to AI training data, the defences that may be available, the litigation and policy landscape, and licensing strategies.
Where copying happens in the training pipeline
Copyright risk can arise at several stages, and each needs a separate analysis:
- Collection: scraping or downloading works to build a dataset creates copies. Circumventing paywalls or other technological protection measures raises separate liability under the Act's anti-circumvention provisions.
- Storage and preprocessing: cleaning, tokenizing and storing datasets involves further reproductions, sometimes for long periods.
- Training: model training processes copies; whether the trained model itself contains a reproduction of a substantial part of any work is debated and may depend on whether the model can regurgitate training material.
- Outputs: outputs that reproduce a substantial part of a training work can infringe independently; see our insight on AI-generated works.
- Distribution: distributing datasets or models to others may raise secondary infringement issues under s. 27(2) where the distributor knows or should know they contain infringing copies.
Because Canadian copyright is territorial, the location of each step matters. Training conducted entirely outside Canada may not infringe Canadian copyright, but making infringing datasets or models available in Canada, or reproducing works on Canadian servers, can bring Canadian law into play.
Fair dealing and AI training data
Section 29 provides that fair dealing for the purpose of research, private study, education, parody or satire does not infringe copyright. The Supreme Court of Canada has said fair dealing is a user's right that must be given a large and liberal interpretation.
In CCH Canadian Ltd. v. Law Society of Upper Canada, 2004 SCC 13 (SCC docket 29320), it set out a two-step test: the dealing must be for an allowable purpose, and it must be fair, assessed through factors including the purpose, character and amount of the dealing, alternatives to the dealing, the nature of the work, and the effect of the dealing on the work.
In Society of Composers, Authors and Music Publishers of Canada v. Bell Canada, 2012 SCC 36, the Court held that "research" is not limited to creative or scientific research and can include consumers previewing music before purchase.
| Fairness factor | Arguments for developers | Arguments for rights holders |
|---|---|---|
| Purpose | Research into language, images and patterns; transformative use | Commercial product development; competing services |
| Character | Copies used internally and not redistributed | Massive, systematic copying and long-term retention |
| Amount | Necessary to learn statistical patterns | Whole works copied, often entire catalogues |
| Alternatives | No practical alternative for general-purpose models | Licensed or public-domain datasets exist |
| Nature of the work | Publicly available material | Paywalled, unpublished or highly creative works |
| Effect on the work | Model does not substitute for any single work | Outputs compete with and devalue the works; licensing market harmed |
No Canadian court has yet applied these factors to model training. The outcome is likely to depend heavily on facts: the source of the data, whether it was lawfully accessed, the commercial nature of the model, whether outputs reproduce training works, and the existence of a licensing market. Developers should not assume that fair dealing will cover commercial training at scale.
Temporary reproductions and other exceptions
Section 30.71 provides that it is not an infringement to make a reproduction that forms an essential part of a technological process, whose only purpose is to facilitate a use that is not an infringement, and that exists only for the duration of the process. The exception may cover transient copies made during computation, but it is unlikely to cover datasets that are retained, and its second condition depends on the underlying use being lawful. Other exceptions, such as those for non-commercial user-generated content or for educational institutions, are too narrow to support commercial training, although they may be relevant to academic research projects conducted within their limits.
Litigation and policy in Canada and abroad
Litigation over AI training data is under way in several countries, including proceedings brought by Canadian news publishers in Ontario against an AI developer. As at our review, no Canadian court had ruled on the merits of a training-data infringement claim, and foreign decisions, which apply different statutes and fair use or text and data mining rules, are of limited direct authority here.
The federal government consulted on copyright in the age of generative AI (ISED consultation, submissions closed 15 January 2024), asking whether Canada should adopt a text and data mining exception, require transparency about training data, or create remuneration mechanisms. Businesses should monitor both the litigation and any legislative proposal, and should expect that any new exception, if enacted, may come with conditions such as lawful access to the source material, respect for opt-outs or transparency about the data used.
Keep a dataset register that records the source, licence terms, collection method, date and any opt-out signals for each dataset used in training. If a claim is made, that record is the starting point for any defence, and it allows problematic data to be identified and removed from future training runs.
Remedies and exposure for AI training data claims
The financial exposure from an AI training data claim depends on the remedies available. A successful copyright owner can seek an injunction, damages and an accounting of profits under s. 35, or elect statutory damages under s. 38.1. For commercial infringement, statutory damages range from CAD 500 to CAD 20,000 for all infringements of each work, which can add up quickly when thousands of works are involved, although courts have discretion to reduce awards where the total would be grossly out of proportion to the infringement.
Injunctions are a particular concern for model developers, because a court order restricting the use of a model trained on infringing data could disrupt a product line. Developers should therefore consider whether they could retrain or replace a model if a dataset became unusable.
Class proceedings are another source of exposure. Authors, artists and publishers can combine claims, and provincial class action statutes make it possible to pursue collective remedies. Indemnities from AI providers to customers do not reduce the provider's own exposure and may increase it.
Practical steps for developers
For developers building or fine-tuning models in Canada, a defensible approach to AI training data usually includes: preferring licensed, public-domain and openly licensed sources; respecting technical opt-out signals and website terms that prohibit scraping; avoiding circumvention of paywalls and other technological protection measures; deduplicating and filtering datasets to reduce memorization; testing models for regurgitation of training works and adding output filters; keeping a dataset register and retention schedule; and documenting the research purpose and design choices that support a fair dealing argument. None of these steps guarantees a defence, but together they reduce both the risk of a claim and the potential remedy.
Licensing strategies for AI training data
Licensing reduces uncertainty and is becoming common for high-value content such as news, books, images and code. Key terms in a training data licence include:
- Scope: the permitted uses (pre-training, fine-tuning, retrieval-augmented generation, evaluation), models covered and whether outputs may be commercialized.
- Exclusivity and term: whether the licence is exclusive, and what happens to trained models after the licence ends; models usually cannot "unlearn" data.
- Attribution and output controls: commitments to prevent regurgitation of licensed works, provide attribution or links, and respect opt-outs.
- Payment: lump sums, per-use fees or revenue shares, and audit rights.
- Warranties and indemnities: the licensor's authority to grant rights, especially for third-party and user-contributed content, and indemnities for claims.
- Privacy: datasets containing personal information must comply with privacy laws, which apply independently of copyright.
Royalties paid to non-resident licensors may be subject to Canadian withholding tax, so the payment clause should say who bears it and whether amounts are grossed up. See our insight on IP licensing agreements.
Risk allocation for businesses buying AI tools
Most businesses face AI training data risk as customers rather than developers, and their leverage is greatest before they sign. When procuring AI tools, they should review the provider's statements about training data, the indemnities offered for copyright claims arising from outputs or training, the conditions attached to those indemnities (such as using built-in filters), and the provider's use of the customer's own data to train models. Customer data used to fine-tune a model should be governed by clear contractual limits and confidentiality, including deletion obligations when the contract ends. See our page on technology transactions.
Data governance and privacy overlap
Copyright is not the only constraint on AI training data. Personal information in training sets is subject to federal and provincial privacy laws, which require a lawful basis for collection and use and impose transparency and security obligations.
Website terms of service may prohibit scraping, creating contractual risk. Trade secret and confidentiality obligations may apply to data obtained from business partners. A single governance process that screens each dataset for copyright, privacy, contractual and security issues is more efficient than separate reviews. It should also cover data obtained from affiliates and vendors, since a dataset bought from a third party carries the same copyright questions as one collected directly, and the buyer should obtain warranties about how the vendor collected it. See our pages on artificial intelligence and copyright and trade secrets.
How TRW & Co helps
We advise AI developers, data providers, publishers and enterprise users on AI training data and copyright risk in Canada. We assess training pipelines and datasets, structure licensing deals, draft and negotiate provider and customer terms, design dataset governance and opt-out processes, and help clients respond to claims and demands. To discuss a dataset, a model under development, a licensing negotiation or an AI procurement, contact us. This article is general information, not legal advice.
Frequently asked questions
Is training an AI model on copyrighted works legal in Canada?
It has not been decided. Copying works into datasets is a reproduction that infringes unless an exception applies. Developers may argue fair dealing for research or the temporary reproduction exception, but no Canadian court has applied them to commercial model training, and the outcome will depend on the facts.
Does Canada have a text and data mining exception?
No. As at the Copyright Act consolidation current to 19 October 2026, there is no specific text and data mining exception. The federal government consulted on the issue, with submissions closing on 15 January 2024, but has not enacted one.
How does fair dealing apply to AI training data?
The dealing must be for an allowable purpose, such as research, and must be fair. Courts consider the purpose, character and amount of the dealing, alternatives, the nature of the work and its effect on the work. Commercial, large-scale copying of entire works is harder to justify as fair.
What should a training data licence include?
Permitted uses and models, exclusivity, treatment of trained models after expiry, output controls, payment and audit rights, warranties of authority, indemnities and privacy compliance. Withholding tax may apply to royalties paid to non-resident licensors.
What should businesses check when buying AI tools?
The provider's training data practices, copyright indemnities and their conditions, whether customer data is used for training, confidentiality protections, and compliance with privacy law. Contracts should allocate responsibility for infringing outputs and data misuse.
Sources
General information only, not legal advice. References to statutes, courts and authorities describe published law and practice at the date shown.


