Anthropic's 1.5 billion dollar settlement covers 482,460 pirated books, not the training that used them

The largest copyright recovery in United States history is routinely described as the price of training an artificial intelligence model on books. The court record says it is the price of downloading them from pirate libraries.

On 20 July 2026 Judge Araceli Martinez-Olguin of the United States District Court for the Northern District of California granted final approval to the class action settlement in Bartz v Anthropic, docket 3:24-cv-05417-AMO, and entered judgment. The Authors Alliance, which read the order, described it as 23 pages, and noted that the fee question was the only part of it to receive extended treatment. The Association of American Publishers announced the approval the same day.

The distinction between training and acquisition was drawn by a different judge more than a year earlier. In June 2025, as Publishing Perspectives records, Judge William Alsup held in a mixed summary judgment decision that Anthropic's unauthorised use of copyrighted books to train its Claude system was fair use, but that the company's decision to keep millions of unauthorised downloads for a permanent central research library was likely infringing. Those downloads came from the shadow libraries Library Genesis and Pirate Library Mirror. The settlement followed.

The settlement class is defined by that library, not by the model. It covers 482,460 works drawn from LibGen and PiLiMi. The fund of 1.5 billion dollars is non-reversionary, meaning money does not return to Anthropic if initial distributions leave a balance. The Authors Guild reported that the settlement provides about 3,000 dollars per work, and that the court described this as four times the statutory minimum of 750 dollars for ordinary infringement. Writer Beware, tracking the order, put the figure at roughly 3,127 dollars per work, calculated across the works actually claimed, and noted that the settlement terms require authors and publishers to split the per work payout evenly where publishing contracts are still in force and both parties claim.

The claims rate was extraordinary and the sources do not agree on it. The Authors Guild reported on 17 April 2026 that claims had been filed for 440,490 of the 482,460 eligible works, a rate of 91.3 per cent, up from 54 per cent as at 19 March 2026, eleven days before the 30 March filing deadline. The Association of American Publishers put the rate at 92.77 per cent. Both organisations were describing the same claims process. Neither statement explains the gap.

The judge was markedly less generous to the lawyers than to the class. Class counsel initially sought 300 million dollars, then reduced the request to 187.5 million, or 12.5 per cent of the fund. The court awarded approximately 101.5 million dollars, about 6.8 per cent, which the Authors Alliance calculated as roughly 2,953 dollars per hour across the 34,381.6 hours of attorney time the court credited. Service awards to the named plaintiffs were cut from the 50,000 dollars each requested to 15,000.

Objections were overruled. According to LexSummary's account of the order, only 350 timely opt-outs covering 1,802 works were submitted. Writer Beware reported that late opt-out requests were mostly refused, with two exceptions granted for authors who are plaintiffs in a separate action. The Authors Alliance noted that the most substantial objections came from people outside the settlement looking in, that the court declined to entertain objections from non-class members, and that it reminded them they remain free to bring their own claims.

What the class gave up is correspondingly narrow. LexSummary records that the release covers claims tied to Anthropic's past acquisition and copying of works on the settlement list, and does not release claims concerning AI outputs or conduct occurring on or after 25 August 2025. The Authors Guild adds that Anthropic must destroy all original files of works obtained from Library Genesis and Pirate Library Mirror, along with any derivative copies, subject to legal preservation requirements, and that the company has said neither dataset was included in its commercially released language models. That last point is Anthropic's account of its own systems.

No money has moved yet, and the sources differ on how fast it will. The Authors Alliance, reading the settlement agreement, says payments begin one business day after the effective date, which it defines as the date on which the time to appeal the final judgment expires, generally 30 days in the Ninth Circuit. Writer Beware quotes the same agreement as giving the settlement administrator 28 days from the effective date to send payments, and defines the effective date as the later of the expiry of the appeal window or the resolution of any appeal. Both report that the window for appeals ran to 19 August 2026. Writer Beware reported that Anthropic pays in instalments through September 2027, with roughly 70 per cent going out in a first distribution and the remainder in a second.

The boundary the case drew is already being cited against AI developers abroad, and not always in their favour. When the Regional Court of Munich I ruled for GEMA against the music generator Suno on 31 July 2026, it expressly distinguished Bartz and Kadrey on the ground that in those American cases the training data had not been substantially made accessible to users in the outputs. On the Munich court's facts, it had been, and fair use failed.