A federal court preliminarily approves Anthropic's $1.5 billion settlement over pirated training books, about $3,000 a work

2025-09-25Copyright

Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal. Sept. 25, 2025) (Alsup, J.) (minute entry granting preliminary approval of class settlement, Dkt. 427), on the Motion for Preliminary Approval of Class Settlement (Dkt. 363, filed Sept. 5, 2025)

After Judge William Alsup's June 23, 2025 order held that training on books was fair use but that pirated library copies were not, the court on July 17, 2025 certified a "LibGen & PiLiMi Pirated Books Class," and trial was set for December 1, 2025. On September 5 the class moved for preliminary approval of a settlement. Anthropic will pay at least $1.5 billion plus interest into a non-reversionary fund; with around 500,000 works in the class, class counsel estimated a gross recovery of $3,000 per work, and Anthropic will pay a further $3,000 for each work added above 500,000. Anthropic will destroy the original files it obtained from Library Genesis and the Pirate Library Mirror and copies originating from them, subject to preservation obligations. The release covers only past conduct through August 25, 2025, and releases no claim, past or future, arising from allegedly infringing outputs of Anthropic's models. Class counsel called it the largest publicly reported copyright recovery and compared the per-work figure with the $750 statutory minimum and the $200 floor for innocent infringement. Funds are allocated equally per work. On September 25, after a September 8 hearing and written questions from the court, Judge Alsup granted preliminary approval of the class settlement, the plan of distribution and the plan for notice.

What it changesPirated training data now has a price per work, and a model budget that does not carry acquisition or licensing cost for its corpus is missing a line. For the patent program the consequence is in diligence and continuity. A patent on a model, a training method or a product built on the model does nothing to offset liability for the corpus the model learned from, so any acquisition of a company, a portfolio or a model license should ask for the acquisition record of the training data alongside the patent schedule. The destruction term shows that a settlement can take datasets out of service, and a product line whose patented features depend on a model trained on those datasets needs a retraining plan. Output claims were expressly left open, so the filtering and attribution mechanisms that keep a model from reproducing its training text remain worth both engineering investment and patent filings.