Anthropic Settles $1.5 Billion Over Covert Book-Scanning Operation 'Project Panama'
Key Takeaways
- ▸Anthropic's covert 'Project Panama' operation purchased and destructively scanned millions of books for Claude training data, with internal documents showing the company deliberately concealed the initiative.
- ▸A federal court ruled that destructively scanning legally purchased books qualifies as fair use, but downloads from piracy sites like LibGen do not—resulting in the $1.5 billion settlement in September 2025.
- ▸The $1.5 billion settlement is the largest known copyright payout in U.S. history, demonstrating the industry-wide shift from web-harvested to book-sourced training data and the high value companies place on high-quality text.
Summary
A Washington Post investigation in January 2026 revealed that Anthropic ran a covert operation called 'Project Panama' starting in early 2024, in which the company spent tens of millions of dollars purchasing millions of print books from libraries, secondhand dealers, and bookstores. The books were then destructively scanned using hydraulic cutting machines and high-speed scanners—a faster, cheaper method that leaves no intact physical copies—with the resulting text fed into the language models behind Claude. An internal planning document stated: 'We don't want it to be known that we are working on this.'
The operation became public through court documents unsealed in a class-action lawsuit brought by authors. In June 2025, federal judge William Alsup ruled that destructively scanning legally purchased books constituted fair use under copyright law. However, court records revealed that in parallel with the book purchases, Anthropic had also downloaded millions of books from piracy libraries such as LibGen—a practice not protected by the fair use ruling. In September 2025, Anthropic settled the case for $1.5 billion, approximately $3,000 per affected work for roughly 500,000 books, making it the largest publicly known copyright settlement in U.S. history.
The discovery reveals broader industry trends. German antiquarian bookstores began reporting unusual bulk orders in 2026 from what appears to be another company acquiring books for AI training. The phenomenon underscores a critical reality: the open internet has been largely exhausted as a training data source, forcing AI companies to pursue high-quality printed text and specialized literature to improve their models—despite the substantial legal and ethical complications.
