On July 20, the U.S. District Court for the Northern District of California, in Bartz v. Anthropic PBC (No. 4:24-cv-05417), granted final approval to a $1.5 billion settlement resolving authors’ claims that Anthropic built its training library out of pirated books — the largest copyright settlement on record.
The settlement resolves only one piece of the underlying case, and it is important to be precise here about which part. Last June, on summary judgment, Judge William Alsup split the case along a line that is likely to define AI training litigation for years: both how a work was acquired, as well as how it was used. Applying the four-factor fair use test, Alsup found Anthropic’s training use “spectacularly” transformative, analogizing it to how a person reads existing books before writing their own: Anthropic’s models, he wrote, trained on the works “not to race ahead and replicate or supplant them — but to turn a hard corner and create something different.” The second factor, nature of the works, weighed against Anthropic, but the third factor was neutral: ingesting entire works was necessary for AI models to learn. And the fourth factor, the amount used and market effect, favored Anthropic since training did not generate outputs that reproduced substantial portions of any book or displace demand for the originals. On that reasoning, training Claude on books Anthropic had acquired legally, through purchase and digitization, was fair use.
But he denied summary judgment on whether downloading millions of books from pirate repositories LibGen and PiLiMi to build a permanent “central library” was itself infringing, reasoning that retaining pirated copies for general-purpose use beyond immediate training needs was its own act, and not a transformative one. That piracy claim, not the fair use question, is what was certified as a class and settled.
The certified class covers roughly 500,000 titles from the more than seven million books allegedly downloaded. Each claiming rightsholder is set to receive about $3,000 per work. Class counsel sought $187.5 million in fees; the court awarded $101.56 million, withholding 10% pending review.
The approval was contested, though. Objectors argued that $3,000 is a fraction of the $150,000 statutory ceiling per willful infringement, and flagged claims-process and notice problems, including omission of Alsup’s fair use order. The court’s questioning at the hearing focused mainly on fees.
There are two takeaways for AI litigation watchers. Bartz is about sourcing, not the legality of training itself — Alsup’s fair use holding remains untested on appeal and unresolved for every other AI developer. And the settlement sets a valuation benchmark: a piracy-based per-work figure that the recently filed Meta, Google, Udio, and Midjourney suits will likely measure claims against.