A federal judge in San Francisco declined to grant final approval to Anthropic's $1.5bn settlement with authors who alleged the company trained its Claude models on pirated books from shadow libraries LibGen and Pirate Library Mirror. Judge Araceli Martínez-Olguín requested more detail on counsel fees (trimmed from 15% to 12.5%), $3m in expenses, an $18.22m cost reserve, and $50,000 service awards for each of the three lead plaintiffs before signing off. The settlement would cover roughly 480,000 works at ~$3,000 per work, with 92% of eligible works already claimed. Objections include claims the settlement is too small, inadequate notice for non-English speakers, and unfair treatment of authors with group copyright registrations. The case stems from a prior ruling that training on pirated copies was not fair use, which pushed Anthropic to settle despite disagreeing with the ruling.
