AI training data and web scraping in Canada: what the OpenAI and Reddit cases mean for founders

Two lawsuits are shaping how founders should think about training data. Neither has decided whether training an AI model on someone else’s content is lawful, but together they show the risk comes from more than one direction.

The Canadian case against OpenAI

A group of Canadian news organizations sued OpenAI in Ontario, alleging their articles were copied without permission to train ChatGPT. OpenAI challenged whether an Ontario court could hear the case. In late 2025 the Ontario Superior Court of Justice held that it has jurisdiction over the claims against several of the OpenAI companies, while dismissing the claims against others (Toronto Star Newspapers Limited v. OpenAI Inc., 2025 ONSC 6217). OpenAI has reportedly appealed. The court hasn’t decided whether training on the articles infringes copyright.

Reddit v. Anthropic

Reddit sued Anthropic in California state court in 2025, alleging its content was collected and used for AI training in breach of Reddit’s user agreement. Anthropic moved the case to federal court, arguing it was really a copyright claim. In March 2026 a federal judge sent it back to state court, finding that the claims are about contractual limits on access, trespass, user privacy and misrepresentation rather than rights equivalent to copyright (Courthouse News, March 2026). What matters for founders is the theory: the claim is built on contract and terms of use rather than copyright. The theory is that a platform that doesn’t own the copyright in user posts may still have a claim if its terms restrict how content can be collected. No court has yet decided whether it succeeds.

Is scraping legal in Canada?

There’s no single rule. Collecting publicly available information isn’t automatically unlawful, but several laws can apply at once:

• Copyright. Copying protected works can infringe unless an exception applies. Fair dealing covers listed purposes such as research and education, but no Canadian court has decided how it applies to AI training. The federal government consulted on copyright and generative AI (submissions closed January 15, 2024) and, as of this writing, hasn’t changed the law.
• Contract. If a site’s terms prohibit scraping and you’re bound by them, scraping can be a breach of contract.
• Privacy. Scraped content often includes personal information. Canada’s Privacy Commissioner and other privacy authorities have said in a joint statement on data scraping that publicly accessible personal information is still subject to privacy law. Quebec’s private sector law adds its own requirements.

What AI startups should do

• Keep a data inventory: where each dataset came from, under what licence or terms, and when.
• Prefer licensed, open-licensed or first-party data, and keep the licences.
• Respect robots.txt and site terms, and stop collecting when asked.
• Filter personal information out of training data, or confirm you have a lawful basis for it. The privacy commissioners’ generative AI principles encourage using de-identified or synthetic data where possible.
• Check what your model providers promise. Indemnities and data-use terms differ widely.
• Expect enterprise customers to ask about this in security reviews and SaaS agreements, and have honest answers ready.

Ownership of what you build on top of that data is its own question. See who owns code a contractor writes.

What platforms should put in their terms

If you host content, including user-generated content, your terms should deal with AI directly: whether automated collection is allowed, whether content may be used to train models, what licence you take from users, and whether you use customer data to improve your own AI features. That last point belongs in your privacy policy too.

Clear terms help most when users actively agree to them, and they work best alongside technical measures and a consistent enforcement record. See AI product terms and privacy and terms.