AI Companies Turn to Pre-2022 Printed Books for Training Data

AI labs are acquiring printed books in bulk through secondary-market platforms such as Better World Books, Alibris and Biblio. According to a bookseller who began noticing a “historic spike” in sales in April, weekly orders have risen from roughly 20 books to “hundreds of books a week.” ISBNdb, which markets itself as the world’s largest book database, offers high-volume acquisition services ranging from 1,000 to 1 million titles per order and guarantees that buyers’ identities remain confidential via strict non-disclosure agreements. Anthropic’s internal documents, revealed in a copyright lawsuit, show the company contracted a scanning firm (Datamation) to digitize millions of purchased copies, suggesting that the bulk purchases are intended for AI model training.

404 Media

The concern is that training on text that already contains AI-generated output can cause “model collapse,” degrading future models. Moreover, authors can deliberately poison AI training data by inserting sabotaging text; printed works from the pre-LLM era lack such modern “poisoning tools,” making them attractive to developers seeking clean, peer-reviewed knowledge.

404 Media

U.S. District Judge William Alsup ruled that Anthropic’s practice of scanning and destroying the physical copies qualified as “clearly transformative” fair use under Section 107 of the Copyright Act.

404 Media

The same bookseller who reported the sales surge expressed mixed feelings: the bulk sales provide financial relief and clear out inventory, yet the destruction of rare, foreign-language, low-circulation titles could make those works harder to obtain for libraries and scholars. He noted that the large orders consist of books with ISBNs, while his stock of rare, non-ISBN volumes has not been targeted.

404 Media

• “Here, every purchased print copy was copied in order to save storage space and to enable searchability as a digital copy,” — William Alsup

404 Media