AI Companies Purchase and Destroy Second‑Hand Books for Training Data

Technology firms in Silicon Valley are buying used books from second‑hand retailers worldwide through intermediary companies. The books are scanned to create large text datasets that train advanced language models such as OpenAI’s ChatGPT, Meta’s Llama, and Anthropic’s Claude. After scanning, the volumes are destroyed, often by cutting their spines, to prevent further distribution.

The practice was disclosed in a Washington Post investigation that identified Anthropic’s internal “Project Panama,” a secret effort to acquire and eradicate physical books for AI training. Librarians in Spain, including a bookseller in Badalona, have reported a sharp rise in bulk purchases of out‑of‑print titles, raising concerns that valuable bibliographic heritage could be lost.

Stakeholders note the shift from previously free online sources to physical copies was driven by copyright lawsuits, prompting firms to seek older, likely out‑of‑rights books that are easier to acquire without legal risk. The trend highlights ongoing debates over copyright, cultural preservation, and the ethics of using printed works as AI training material.