Judge Approves Anthropic's $1.5 Billion Copyright Settlement

The US District Court for the Northern District of California granted final approval on July 20 to the $1.5 billion settlement in *Bartz v. Anthropic*. The court found the agreement fair, reasonable, and adequate after reviewing the parties' filings, a May 14 hearing, and submissions from 54 class members and putative class members.

The certified class covers eligible copyright owners of books on the settlement's Works List that appeared in versions of LibGen or PiLiMi downloaded by Anthropic. The order says 440,490 of 482,460 listed works, or 91.3%, had been claimed as of April 16. It describes an estimated payment of about $3,000 per work, before fees and costs, while noting that actual distributions depend on the approved allocation process and competing ownership claims.

The settlement resolves covered claims tied to Anthropic's past acquisition and copying of the listed works. The order states that the release does not cover claims about past AI outputs or future conduct on or after August 25, 2025. The court also overruled objections to the settlement and disallowed most late opt-out requests.

The judge awarded class counsel about $101.6 million in fees, substantially below the requested $187.5 million, plus approved expenses and a future cost reserve. The court retained jurisdiction over implementation and distribution disputes.

A separate court order in the case treated training on the authors' works as fair use while leaving acquisition-related claims in play. That distinction is the operational lesson for AI practitioners: a model-development workflow can face different legal analysis at different stages. Training methodology, dataset acquisition, storage, lineage, licensing, and deletion controls are related, but they are not interchangeable.

For teams building or buying foundation-model capabilities, the settlement is a concrete reason to maintain source inventories, rights metadata, acquisition records, and retention policies alongside model cards and evaluation results. Those controls do not decide legal liability, but they make it possible to show what entered a corpus, on what basis, and what happened to it later. The ruling closes this class settlement; it does not settle the broader copyright questions around generative AI.