As a result of the $1.5 billion settlement in the Bartz v Anthropic class action suit, rightsholders (book authors and publishers) who registered their works—and that’s about 500,000 books out of a reputed 7 million that Anthropic downloaded—can now receive about $3,000 per eligible work. Most payouts will be split, with half going to the author and the other half to the publisher. U.S. federal judge Araceli Martinez-Olguin approved the settlement gov.uscourts.cand.434709.680.0_5.pdf on 20 July 2026. In addition to cash disbursements, Anthropic is ordered to destroy the internal unauthorized copies of the books and any pirated files it downloaded.
If a librarian published one book that qualifies to receive a payout from Anthropic, the amount will be around $1,500. The publisher gets the other $1,500. A publisher such as Bloomsbury, https://bloomsbury-ir.co.uk/media/press_releases/2026/220726.asp which had 14,087 titles qualify, stands to receive $21,130,500. Now that’s a nice chunk of change.
The books in question were pirated from LibGen and PiLiMi. To qualify for inclusion in the class action, books had to be registered with the U.S. Copyright Office and have ASBN or ASIN numbers. It’s too late to submit a claim, however. According to the Anthropic Copyright Settlement website www.anthropiccopyrightsettlement.com/dates the key dates to submit a claim have passed. If authors did nothing, the outcome "will depend on what the other rightsholders for your works do". Most likely, that means no money will be forthcoming.
The settlement is a victory for these particular copyright owners. What it is not is a clearcut definition of what comprises copyright infringement. The original judge, now-retired William Alsup, included in his settlement ruling the notion that training language models on books is fair use. Where Anthropic went wrong was using pirated versions of the books it used and keeping them in what resembled an internal library that could be used for purposes other than training.
Keep in mind that this is a settlement rather than a decided copyright court case ruling. Its legal precedent is limited. And there’s no shortage of pending copyright cases worldwide concerning unauthorized use of copyrighted materials to train language models. The concept of transformative use is widely used to justify this use of copyrighted materials, but copyright law can vary from country to country. What is clear is that fair use does not extend to materials acquired unlawfully. No illegal repositories qualify as fair use training for AI purposes. That’s infringement. As for the rest? We’ll have to wait for the courts to decide.