The copyright fight, as a fight

Three questions that get run together

Every argument about AI and copyright is actually three, and they have different answers.

Was the copying at training lawful? Building a training set means making copies of millions of works. Is that infringement, or an exception?

Is the output lawful? When a model reproduces a passage, a lyric or a recognisable image, is that a copy? Most parties on both sides agree it can be; the labs filter for it.

Does the market matter? Even if no output copies any work, does a model that competes with the works it learned from harm their market in a way the law should count?

Hold those apart and the positions become readable. This lesson gives the shape of the disagreement in five places. It does not resolve it, because no court or legislature has, and it is not legal advice.

The United States: fair use, case by case

American law allows “fair use” of a work without permission when the use is transformative and does not substitute for the original, judged case by case. The labs’ position is that training is reading at scale, and transformative. The publishers’ position is that it is copying at scale and that a market for licensing exists and is being destroyed.

In June 2025 two federal judges in California ruled within days of each other, on cases brought by authors against two labs. Both found that training on the books was fair use on the records before them: one judge called it “exceedingly transformative”. Both drew lines. One held that obtaining the books from a pirate library was not fair use at all, sent that question toward trial, and the lab settled for around one and a half billion dollars, roughly three thousand dollars a book. The other wrote that a better-argued case might win on the third question (that a flood of competing works could dilute the market even without copying) and that the authors had not made that argument.

Meanwhile the largest case, brought by a newspaper against the biggest lab in late 2023, was proceeding, and a separate 2025 ruling found that a non-generative legal-research tool trained on a rival’s material was not fair use. The American answer, so far: two trial courts have found training on books to be fair use on the records before them, one of them even though the books came from shadow libraries. One court held that keeping a permanent library of pirated copies was not fair use. The market-dilution question is open. And at the end of September 2026 the first appeals court to rule on AI training, the Third Circuit, upheld the finding that training the legal-research tool on its rival’s material was not fair use.

The United Kingdom: an exception that was not made

British law permits text and data mining only for non-commercial research. In late 2024 the government proposed a commercial exception with a right for creators to opt out. The proposal met the most concerted opposition the creative industries had mounted in years. The Data (Use and Access) Act 2025 passed in June 2025 only after the Commons had repeatedly stripped out Lords amendments that would have made AI developers disclose their training material. Instead, the Act required the government to report back within nine months. Its report, published on 18 March 2026, set the opt-out exception aside and proposed no immediate change to the law. And the leading British case, a photo agency against an image-model company, produced a judgment in late 2025 that went largely the company’s way on the specific facts (the training had happened outside the UK, and the court found the model file itself was not an infringing copy because it did not store the works) while leaving the training question itself largely unanswered.

The European Union: an exception with an opt-out

The EU’s 2019 rules allow commercial text and data mining unless the rights-holder has opted out in a machine-readable way. The AI Act makes frontier labs honour those opt-outs and publish a training-data summary. A German court found in 2024 that assembling a public research dataset fell under the research exception, and a Munich court found in late 2025 that a model reproducing song lyrics it had memorised infringed. The European answer, so far: mining is permitted if you were not told no, and reproducing is not.

Japan: the widest exception

Japan’s 2018 law permits use of works for analysis, including commercial machine learning, unless it unreasonably harms the rights-holder’s interests. It was called a paradise for training and is now being reconsidered under pressure from creators.

India: no exception, and a live case

India’s law has no text-and-data-mining exception, and its fair-dealing provision is a closed list (research, private use, criticism) that training does not obviously fit. A news agency sued the largest lab in the Delhi High Court in late 2024; the lab argued that its servers and training were outside India. The case was proceeding, the court had appointed advisers, and a government committee was examining whether to create a licensing scheme. India’s answer, so far: none, and the first one is being written.

What the fight is about

Strip the law away and the disagreement is about who pays whom. If training is lawful without a licence, the value of the works flows to the labs. If it requires one, a market forms and the largest rights-holders are paid, as the music labels began to be through settlements. Neither outcome is obviously the just one for the individual writer or illustrator, whose share of any licence is small and whose work was in the pile either way. The next lesson turns the question round: who owns what the model makes.