Anthropic illegally trained Claude with Stanford professors' books

A federal judge’s ruling in the Bartz v. Anthropic case drew a distinction between acquiring books legally to train Claude and building a library from pirated copies. The judge found Anthropic could not use illegally obtained books to create a permanent central collection, while treating training on lawfully acquired books as fair use. A $1.5 billion settlement covers authors and publishers of nearly 500,000 registered works, but Stanford professor Tobias Wolff and 27 other authors say its compensation is inadequate. Wolff says 10 of his publications were used without permission and warns that uncompensated AI training threatens writers’ livelihoods, especially those early in their careers.

Wolff and English professor Chang-rae Lee opted out of the settlement and joined a separate lawsuit seeking statutory damages of up to $150,000 per infringed work, rather than the roughly $3,000 per work offered through the settlement. The plaintiffs allege Anthropic downloaded and distributed books from shadow libraries and stored pirated files centrally; success could bar Claude training on their works. Stanford intellectual-property scholar Paul Goldstein points to the practical difficulty of licensing millions of works, while also stressing authors’ interest in attribution. The disputes form part of a wider wave of more than 70 copyright suits against AI companies. Authors’ advocates are pursuing litigation and legislation to require licensing, compensation, and controls on how copyrighted works appear in AI outputs.