Is it legal for AI companies to train on books, news and music?

To learn to write, a language model is trained on enormous amounts of text: it reads billions of sentences and adjusts its internal settings to predict the next word. Image and music models learn the same way from pictures and recordings. Much of that material is protected by copyright, and the companies usually did not ask permission. Whether that is legal is one of the biggest open questions in AI, and as of October 2026 it still has no single answer.

In the United States, where most of the cases are, the key defense is fair use: a rule that allows some copying without permission, judged on four factors. Courts look at the purpose of the use (and whether it is "transformative", meaning it does something new rather than substituting for the original), the nature of the work, how much was copied, and the effect on the market for the original. Each case turns on its facts, which is why results so far point in different directions.

Three rulings, three different answers

The clearest way to see the pattern is to put the main decisions side by side.

Read across the rows and two questions keep coming back: where did the copies come from, and does the AI compete with the original? The Third Circuit, a federal appeals court, stressed that ROSS's system could not generate original expression and was built as a commercial substitute for Westlaw. That distinguishes it from chatbots like Claude or ChatGPT, but it shows that "it's AI" is not a defense in itself; the court called it an ordinary copyright case.

News and music are less settled. The Third Circuit's opinion mentions a pending case in New York against OpenAI, in which the US Justice Department filed a statement in September 2026 arguing that training a large language model is transformative. A government statement is an argument, not a ruling.

A worked example: checking a claim you read online

Headlines often say "court rules AI training is legal" or "illegal." You can test a claim like that with a few questions, and an AI assistant can help you organize the answers from the decision itself:

I will paste a summary of a court decision about training AI on copyrighted material. Answer in a short table: 1) which court and which country, 2) what material was copied and how it was obtained, 3) whether the AI generates new content or competes directly with the original, 4) what the court actually decided, and 5) what the court said it was NOT deciding. Quote the decision's own words for point 5. If the text does not answer a question, write "not stated" instead of guessing.

Point 5 matters most. Judge Chhabria's ruling in the Meta case says plainly that it does not establish that Meta's training is lawful; it only finds that these authors made the wrong arguments.

Why did Anthropic pay $1.5 billion while Meta won its fair-use ruling?

Because the two companies were judged on different questions, and Anthropic's case reached a question Meta's did not.

In Bartz, the court ruled in June 2025 that using books to train Claude was fair use. But Anthropic had also downloaded millions of books from the pirate sites LibGen and PiLiMi to build a central library, and the court sent the question of those copies to trial, with damages under US law ranging from $200 to $150,000 per work. Faced with a December 2025 trial, the parties settled. The settlement creates a $1.5 billion fund, about $3,000 per eligible work before fees and costs, which the court-approved notice describes as the biggest settlement ever in a copyright class action. Anthropic must destroy the downloaded files, and it stated that those datasets were not in the training data of its commercially released models. The court gave final approval on July 20, 2026, and the first payments are expected on or before November 15, 2026.

The settlement covers past claims only. It does not cover claims about what AI models produce, and it is explicitly not a license to train on copyrighted works. In other words, it pays for how the books were obtained, not for permission to learn from them.

In Kadrey, Meta had also used shadow libraries, but the summary judgment ruling of June 2025 addressed only whether training on these 13 authors' books was fair use. The judge found the authors had not shown market harm and rejected their two main theories. The judge suggested that the argument with the best chance, market dilution (AI output flooding the market with competing works), had barely been raised. A separate claim about distributing files while torrenting was left open. Meta won on the record in front of the court, not on a principle that training on copyrighted books is always allowed.

This article explains US cases in general terms and is not legal advice; courts in other countries, including Canada, apply their own copyright rules.