Anthropic hit with $1.5B copyright settlement; court affirms AI training on books is fair use

A U.S. federal judge granted final approval for Anthropic to pay $1.5 billion to settle a class-action copyright lawsuit filed by authors. The landmark ruling affirms that training AI models on published books is legal fair use under copyright law, but the court found that Anthropic's operation of a central library of 7 million pirated books violated authors' and publishers' copyrights—the infringing part was the storage and distribution infrastructure, not the training itself.

The settlement averages roughly $200 per title across all involved works and authors. Over 91% of authors and publishers covered by the settlement have claimed their share. The ruling opens the door for other AI companies—including Meta (which torrented 82TB of pirated content), NVIDIA (which used automated scripts to download books), and others facing similar lawsuits—to argue that training itself is legal, even if they face separate liability for how they sourced the data.

For architects building on foundation models, this sets an important precedent: AI training on existing works is not copyright infringement per se, but the mechanics of obtaining that data can be. Anthropic still faces other disputes, and some groups have opted out of this settlement to pursue separate claims, but the signal is clear that the courts are distinguishing between training methodology and data acquisition.