U.S. judge says AI training legal but fines Anthropic $1.5 billion for illicit book copies

Federal judge William Alsup ordered Anthropic to pay $1.5 billion to a group of authors whose books were used to train the company’s language models. Alsup found the training itself lawful and imposed the fine because Anthropic obtained the books from illegal shadow libraries on the internet, not because the learning process infringed copyright. “Like any reader who aspires to be a writer, Anthropic’s models learned from works not to replicate or replace them, but to point in a different direction and create something different,” the judge wrote.

Intellectual-property and technology lawyer Cathy Gellis said the ruling is convenient for AI companies. She noted that $1.5 billion is a small amount for a firm that expects $200 billion in annual revenue by 2028. “This is good news for AI training that looks at what happens and thinks it’s akin to reading a protected work, not copying it,” Gellis said. She added that copyright law concerns copying, not the experience, consumption or reading of a work.

The decision also underscores that U.S. copyright law has not been updated since 1976, forcing judges to interpret five-decade-old guidelines in cases that will shape the industry’s future. “Everyone is now worried because the law is everywhere, and that’s because of this question,” senior lawyer and founder of the IP and Media practice at JWL International Jason Henderson said. “They know the model was trained on massive amounts of data, and the law hasn’t really bridged the gap.”

The dispute pivots on the fair-use doctrine, which permits use of protected material without explicit permission for criticism, parody, education and similar purposes. Courts weigh the purpose of the use, the amount taken, and the effect on the market. “Copyrights always deal with protecting and growing the market,” Henderson observed. “Courts are scattered in their reasoning in AI cases. What tends to win: if you train on someone’s property with the purpose of directly competing, courts will frown. If it is not going to compete, they tend to find ways to approve it.”

The Thomson Reuters v. Ross Intelligence case is frequently cited as a precedent. In that lawsuit, copying content for a competing AI-driven legal research platform was ruled non-transformative and therefore unlawful.