Is training AI on copyrighted books legal? Courts are still deciding

Most published authors have, without their knowledge or consent, had their books used to train the large language models behind chatbots like ChatGPT, Gemini, and Claude. Whether that is legal, though, has no clean answer yet. “It’s very complex and there are a lot of raw feelings about what is happening, both for and against,” Cathy Gellis, an attorney specializing in intellectual property and technology law, told TechCrunch.

Last year, Judge William Alsup ordered Anthropic to pay a $1.5 billion copyright settlement to a group of writers whose works trained the company’s models. But Alsup separately ruled that the AI training itself was lawful — the penalty was for pirating the books from illegal shadow libraries, not for training on them. Alsup compared an LLM ingesting text to a writer studying literature, writing that Anthropic’s models trained on the works “not to race ahead and replicate or supplant them — but to turn a hard corner and create something different.” Gellis called the ruling favorable for AI companies, noting a $1.5 billion fine is comparatively small next to a company projecting about $200 billion in annual revenue by 2028.

The legal fight centers on fair use, the copyright carve-out that permits using protected works without permission for purposes like criticism, education, or transformation. Courts weigh factors including the purpose of the use and its effect on the original work’s market — and, according to attorney Jason Henderson, they have leaned toward allowing AI training when it doesn’t directly compete with the copyrighted material’s original market, and against it when it does. Henderson pointed to Thomson Reuters’ suit against Ross Intelligence, which used Reuters’ content to build a competing legal-research platform: Judge Stephanos Bibas ruled that use was not fair, because Ross’s product lacked “a ‘further purpose or different character’” from Reuters’. Authors have not yet succeeded in arguing that AI chatbots compete with them in a similar way.

Copyright law hasn’t been updated since 1976, leaving judges to apply decades-old standards to AI training disputes that could shape the industry’s future. Separately, courts have found that fully AI-generated works aren’t copyrightable at all, under the ruling in Thaler v. Perlmutter — a distinct question from whether training on copyrighted material is legal, but one that raises its own unresolved issues about how to determine what portion of a work was AI-made. With most major AI companies still facing pending litigation over these questions, Gellis said, early rulings are shaping industry behavior even though later cases could still overturn them.

Why it matters

Until higher courts or Congress settle the fair-use question for AI training, companies building foundation models are operating under law that could still shift beneath them — and the difference between “trained on it” and “pirated it” is turning out to matter more than whether training itself is permitted.