The copyright question nobody has settled

Why this lesson refuses to give you an answer

This is not legal advice, and no honest lesson on this subject in 2026 can end with a rule you follow. The law differs by country, several of the important cases are unresolved or under appeal, and legislatures are actively rewriting the ground. What you can have is the shape of the disagreement, which is stable enough to be useful: knowing which question is being argued lets you read a headline correctly and ask a lawyer a precise question instead of a vague one.

There are three separate questions, and most confused arguments are two people answering different ones.

Question one: is training itself an infringement?

Training copies material: into a crawl, into a filtered corpus, into GPU memory. Copying is the act copyright regulates. The dispute is whether this particular copying is permitted.

United States. The argument runs through fair use, and specifically through whether training is transformative and what harm it does to the market for the original. Courts have begun to split the question: some rulings have treated the act of training on lawfully obtained works more favourably than the act of assembling a library from pirated sources, which puts the spotlight on acquisition rather than training. That distinction is doing a lot of work and is still being tested.

European Union. The DSM Directive created text and data mining exceptions. Research organisations get a broad one. Everyone else gets one that applies only where the rightsholder has not reserved their rights in a machine-readable way, which is what turns robots.txt and similar signals into something with legal weight in the EU that they do not have elsewhere.

Japan. Article 30-4 of its copyright act permits use of works for machine analysis in fairly broad terms, with limits where the use unreasonably prejudices the copyright owner. It is generally read as the most permissive of the major regimes.

United Kingdom. A narrow non-commercial research exception, repeated consultations on a broader one, and no settled position at the time of writing.

India. Fair dealing under section 52 is an enumerated list of purposes rather than an open standard like US fair use, which makes the American arguments non-transferable. This matters if you are building here.

Question two: what about the output?

Separately from how the model was trained, a specific generation can reproduce protected expression. Research has repeatedly shown that models emit training text verbatim under some conditions, and that the rate rises with duplication in the corpus and with model size. Where an output is substantially similar to a protected work, that is an ordinary infringement question and the training argument does not help you.

This is the question you actually control. Deduplicated training data reduces it. So does not prompting a model to reproduce a specific work, and so does checking output against sources when the stakes are high.

Question three: can the output be protected?

Whether what the model produces is itself copyrightable is a third, independent question. The US Copyright Office has taken the position that human authorship is required, registering works where a human contributed enough and refusing purely machine-generated material. Other jurisdictions take different views; the UK has an unusual provision for computer-generated works. If your business depends on owning what you generate, this is the question to ask about, and it has nothing to do with the first two.

What this means for you, practically

• Your exposure comes from what you do, not from what the lab did. Using a model is a different act from training it. Most of the litigation is aimed at trainers and at people who assembled corpora, not at downstream users.
• Open weights come with no indemnity. Several hosted API vendors offer contractual indemnification for copyright claims arising from output. A file you downloaded offers you nothing of the kind. That is a real and underrated difference between renting and self-hosting, and it belongs in the comparison alongside cost and latency.
• Keep records. Which model, which revision, which licence, which date, what you generated and when. If a question arrives in two years, the record is the difference between an afternoon and a month.
• Treat generated code and generated text differently. Code has a strong culture of licence compliance and tooling to match; a snippet that matches a licensed repository is a familiar category of problem with familiar remedies.

If someone tells you this question is settled, ask them which of the three questions they mean, and in which country. The confident answers on both sides are usually answering a different question from the one you asked.