Anthropic to Pay $1.5 Billion Over Pirated Books as EU Bars AI Training Over Rights Holders' Objections

The settlement in Bartz v. Anthropic draws a crucial distinction in AI copyright law: while training on lawfully acquired books was deemed fair use by a U.S. court, the act of copying pirated digital works directly infringed copyright. This clarifies that AI companies cannot simply use any available data, even if the subsequent AI model transforms the content. The ruling highlights the legal risks associated with sourcing training data without proper licensing or consent.

Anthropic's shift to purchasing physical books in 2024, internally called the "Panama Project," indicates a recognition of these legal challenges and a move towards more legitimate data acquisition. Its engagement with distributors like Woongjin Booksen in South Korea for "destructively scanning" books shows a global effort to secure data, even as South Korea's existing copyright law on distribution rights may not fully address sales to AI companies.

The conflicting court decisions in the U.S., the Anthropic settlement versus the Thomson Reuters win against Ross Intelligence, demonstrate the evolving and inconsistent legal landscape for AI training data. While one case focused on the act of copying pirated material, another upheld copyright claims over the use of summaries for training, suggesting that the nature and source of data, along with the specific use, will continue to be critical factors in future disputes.