Anthropic's $1.5B tab: a court rules on the books Claude was trained on

A federal court in California has signed off on a roughly $1.5 billion settlement between Anthropic, the company behind the Claude chatbot, and a class of authors whose books the company ingested to train its AI. The dollar figure, approved on July 20 by U.S. District Judge Araceli Martinez-Olguin, is the largest of its kind in an AI copyright case. The line the ruling draws is narrower, and more useful, than the headline.

The settlement is not a verdict that training an AI on books is, by itself, infringement. That question was already answered the other way. In 2025, U.S. District Judge William Alsup ruled that training on books Anthropic had lawfully acquired was "exceedingly transformative" and protected by fair use. What Alsup split, and what the settlement now prices, is the means of acquisition. Training Claude on the result of reading a book you own is one thing. Sourcing that result by deliberately buying and scanning books you intend to throw away, and by ingesting two named pirate libraries, is another.

The first half of that line comes from the court record, which described an internal Anthropic initiative to bulk-buy physical books, scan them to feed the training pipeline, and dispose of the originals, a program reporting has named "Project Panama." The second half comes from the same record: Anthropic also ingested material from Library Genesis (LibGen) and the Pirate Library Mirror (PiLiMi), shadow libraries of books no one had licensed. Together, those two channels, the bulk-buy-and-trash program and the pirate libraries, are the conduct the settlement is paying for.

The suit started in 2024, when authors Andrea Bartz, Kirk Wallace Johnson, and Charles Graeber filed a proposed class action on behalf of writers whose works Anthropic had ingested. The case went through two judges. Alsup, who had presided over the early fair-use fight, retired before final approval; Martinez-Olguin took the settlement over the line. The math, as the Authors Guild reported, works out to roughly $3,000 for each book Anthropic illegally acquired, a per-unit price tag that puts a number on what "industrial-scale acquisition" actually costs when a court makes you price it.

Class counsel had asked for fees of 20% of the fund; Martinez-Olguin reduced the award to under 7%, about $101 million. The three lead plaintiffs, who had sought service awards of $50,000 each, were trimmed to $15,000 apiece. The cuts are governance texture, not exoneration: the court found the conduct actionable, the settlement is being paid, and the reduction is the difference between a fee request the judge called disproportionate and one the settlement fund could absorb.

What the ruling does not do is close the line of cases testing the same boundary. Anthropic still faces a suit from Reddit alleging the company continued scraping user data after a stop-and-desist notice, and a separate action from Chicken Soup for the Soul that names Anthropic alongside OpenAI, Meta, Apple, and NVIDIA. Both turn on the same axis the Bartz case did: not whether the model was trained on the data, but how the data was acquired, and what the defendant knew when it acquired it. The Reddit case adds a second fault line, a stop-and-desist receipt, that the Bartz plaintiffs did not have to put in front of the court. If a jury credits it, the per-record price tag climbs.

That is the practical takeaway. "Fair use" is a defense for the model you build. It is not, the Anthropic settlement now confirms, a license for the pipeline you use to build it. The next wave of AI copyright cases will not be arguing about whether training is transformative. They will be arguing about what an AI lab has to know, document, and refuse about its inputs before "transformative" still applies.