Anthropic settles with authors for $1.5B (Bartz case)
2025-09-05 · Publisher : Anthropic · Domain : Artificial Intelligence
Sued for downloading pirated books, Anthropic agrees to pay at least $1.5 billion to authors and publishers, about $3,000 per work: the largest publicly reported copyright settlement in the United States.
What changes
- At least $1.5 billion for about 500,000 works, roughly $3,000 per title before fees.
- In June 2025, Judge Alsup ruled that training on lawfully acquired books was fair use, but not building a library of pirated copies.
- Anthropic must destroy the files obtained from LibGen and PiLiMi; the deal only covers past conduct.
- Preliminary approval on September 25, 2025, final approval on July 20, 2026.
- Training data provenance becomes a major financial risk for model providers.
On September 5, 2025, the parties to the Bartz v. Anthropic class action, filed in 2024 by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, submit a proposed settlement: Anthropic agrees to pay at least $1.5 billion to the rightsholders of books it had downloaded from pirate libraries. ### What happened - **The June 2025 ruling**: Judge William Alsup (Northern District of California) finds that training models on purchased and scanned books is fair use. However, he denies that protection for downloading and keeping pirated copies, taken from LibGen in 2021 and PiLiMi in 2022. - **The risk**: with the class certified, Anthropic faced a trial in which statutory damages per work could reach far higher amounts. - **The deal**: about $3,000 per work for nearly 500,000 titles, paid in several installments, destruction of the pirated files, and no license for the future: the settlement only covers past conduct. ### Key figures | Item | Value | |---|---| | Minimum amount | $1.5 billion | | Works covered | about 500,000 | | Payment per work | about $3,000 before fees | | Preliminary approval | September 25, 2025 | | Final approval | July 20, 2026 | ### Why it matters The case draws a dividing line that concerns the whole sector: training is not unlawful in itself in the United States, but the way data was obtained can prove very expensive. It also sets a benchmark for the value of a work in future licensing negotiations. ### Caveats This is a settlement, not a legal precedent: the fair use ruling comes from a trial court, and other courts may decide differently. The question of model-generated content is not addressed. ### Key takeaways For companies, the traceability of training data becomes a criterion for choosing model providers and a matter of legal risk management.
Links
Tags : IA, Droit d'auteur, Justice, Anthropic, Données