American judges are still deciding whether AI models read other people's books, or copy them - and a whole industry hangs on the answer.
Fair use isn't a hole in the law - it's a built-in safety valve. A judge weighs four factors: what purpose the other work is used for, what its nature is, how much of it was taken, and whether the new thing wrecks the market for the original. If the result is "transformative" enough - it makes something new, not just sells a copy - the court can decide there's no theft, even without the author's permission.
AI companies put exactly this question at the core of their defense. To "learn" what human text sounds like or what a photo looks like, a model has to pass through a huge volume of other people's books, articles, music, images. The companies' argument is simple: the model doesn't store or resell the original - it just learns from it, the way a person reads hundreds of books and then writes their own. Opponents see something else: this isn't reading, it's digesting entire libraries to produce a product that then competes with and displaces the very authors whose work it used without paying a cent.
The first serious court ruling on the topic came against Anthropic - the company behind the Claude chatbot, founded by engineers formerly at OpenAI. The judge split the ruling in two: training the model on legally purchased books is fair use, because it's transformative. But the way the company got hold of part of those books - millions of pirated copies downloaded from illegal libraries online - is not fair use, it's plain theft. The result: a $1.5 billion settlement, payable to the authors of the affected books, about $3,000 per title.
The lawsuit by the New York Times - the American newspaper - against OpenAI, the company behind ChatGPT, is still ongoing and looks much darker for the industry. The paper claims OpenAI is withholding evidence and is pushing the court to impose sanctions. The dispute is whether the chatbot was merely trained on those articles, or sometimes reproduced whole passages of them verbatim - which is far harder to defend in court than simply "learning style". The case still has no verdict, and it's exactly this one that could become the precedent for the whole industry.
In parallel, there's a lawsuit against Meta, the parent company of Facebook and Instagram. The court has already partly dismissed the authors' claims on fair-use grounds for the training itself, but left the accusations of pirated book downloads open. The picture emerging from all these cases is the same: learning from someone else's content is allowed. How the companies got hold of it is a separate, much more dangerous question.