Training AI on Copyrighted Books: Two Courts, Two Contradictory Rulings

In 2025, Judge William Alsup ruled in the Anthropic case: training an AI model on authors' books is legal. He compared it to "a writer studying literature" rather than copying. At the same time, he ordered Anthropic to pay $1.5B because the books had been downloaded from pirate sites — a violation even when the training process itself is lawful.

That same year, a different court reached the opposite conclusion in Thomson Reuters v. Ross Intelligence. Judge Stephanos Bibas ruled that using someone else's content to build a competing product is not fair use. His reasoning: such use does not "transform" the original — it directly competes with it. Two different courts, two active precedents.

The line isn't drawn at training itself, but at the outcome: does the AI undermine the market for the original work? Copyright law is 50 years old — written for books and vinyl records, not AI models. A unified precedent that sets the rules for the entire industry is still a long way off.