Key facts
- AI companies are acquiring and scanning millions of physical books for training data.
- Anthropic's 'Project Panama' involved buying, scanning, and destroying books.
- The practice has led to unusual bulk purchase enquiries at rare bookshops.
- A US federal judge ruled Anthropic's scanning process was protected by fair use.
- Older printed books, particularly those published before 2022, are sought after.
- Companies like ISBNdb are facilitating these bulk book purchases for AI labs.
AI companies are reportedly targeting rare bookshops, including those in London, to acquire physical books for training their large language models. Anthropic, the developer of Claude, is reported to have launched an internal program called 'Project Panama' where it purchased, scanned, and destroyed millions of second-hand books. This practice aims to obtain high-quality training data that predates the proliferation of AI-generated content online, with older books being particularly prized. Booksellers have reported receiving unusual, large-volume enquiries from anonymous buyers, mirroring similar reports across Europe. A US federal judge has ruled that Anthropic's method of buying books, scanning them, and destroying the originals falls under fair use protections. Companies like ISBNdb are now facilitating these bulk purchases, acknowledging the reputational risks associated with the practice, such as the headline 'AI company destroys two million books'.