A judge approved Anthropic's $1.5 billion settlement with book authors
Judge Araceli Martinez-Olguin gave final approval to the $1.5 billion settlement over roughly 482,460 pirated books used to train Claude. Authors are getting about $3,000 per work.
A federal judge signed off on Anthropic's $1.5 billion payout to authors whose books it pirated.
A federal judge gave final approval this week to a $1.5 billion settlement between Anthropic and a class of book authors over pirated training data. The deal covers roughly 482,460 books Anthropic pulled from the piracy sites LibGen and PiLiMi to train Claude. It is one of the largest copyright settlements ever reached, and it does not touch the separate question of whether training an AI model on copyrighted books is fair use.
What the settlement actually covers
Authors and publishers whose books were in the pirated set are getting roughly $3,000 per work on average, and Anthropic has to destroy the pirated files it holds. The case, Bartz v. Anthropic, is led by author Andrea Bartz and was filed in 2024. Judge William Alsup oversaw it for most of that time and rejected an earlier version of the settlement in September 2025 as incomplete. He has since retired, and Judge Araceli Martinez-Olguin, who took over the case, granted final approval in the Northern District of California, in San Francisco.
Why this is not the fair use ruling
Alsup had already ruled, in an earlier and separate decision, that training Claude on the books was transformative fair use. This settlement does not revisit that finding. It resolves a different claim: that Anthropic infringed copyright simply by keeping around 7 million pirated books in a central library, regardless of whether any individual book was ever actually used to train a model. That retention claim, not the training question, is what Anthropic paid to settle.
The numbers behind the payout
The largest known copyright recovery in history.
About 91.3 percent of the roughly 482,000 listed works have already been claimed by their authors or publishers, who are now owed payment. Getting to this point took two years from the original filing, including a full rejection and a renegotiation, so the payout process itself is likely to take a while longer to actually reach everyone who is owed.
The bigger picture for AI labs
The Decoder's read on the ruling is that it hands the AI industry its biggest legal win to date, precisely because the fair use finding survives intact. Anthropic paid $1.5 billion for how it stored pirated copies, not for training a model on copyrighted work in a transformative way. That distinction is what every other AI company facing a similar suit will be pointing to next.
Why a build studio cares
We build on Claude and recommend it to clients, so how its training data was sourced is not an abstract question for us. This settlement puts a real number, about $3,000 per pirated book, on what training data provenance can cost after the fact, and it draws a clear line: transformative use of content you have a legitimate right to is defensible, quietly stockpiling pirated copies regardless of whether you used them is not. Any AI vendor's data sourcing practices are now a diligence question with an actual dollar figure attached, not a hypothetical one.
Next step: read the AP's coverage and The Decoder's analysis of what it means for other AI labs. If data provenance is a live question in your own AI build, write to us at hello@gattyworks.com.