Anthropic agreed to a $1.5bln settlement in a 2024 class action by authors alleging the company downloaded millions of pirated e‑books from LibGen and other shadow libraries and included about 482,000 copyrighted works in Claude’s training data. The court split the legal issues, finding that using lawfully obtained texts for model training can be fair use, but downloading and long‑term storage of works from pirate sites constitutes infringement. The settlement averages roughly $3,000 per work an

Anthropic agreed to a $1.5bln settlement in a 2024 class action by authors alleging the company downloaded millions of pirated e‑books from LibGen and other shadow libraries and included about 482,000 copyrighted works in Claude’s training data. The court split the legal issues, finding that using lawfully obtained texts for model training can be fair use, but downloading and long‑term storage of works from pirate sites constitutes infringement. The settlement averages roughly $3,000 per work and requires destruction of the pirated files. Anthropic had earlier accused some Chinese AI teams of extracting Claude outputs via API for model distillation; the court noted those practices are legally distinct. Core takeaway: model training is not per se illegal, but data acquisition must comply with copyright law.