AI Hunts for Data: Booksellers Receive Mysterious 3000-Book Request

AI companies are increasingly sourcing physical books, moving beyond the internet for training data. Antiquarian booksellers in the Netherlands reported receiving an unusual email requesting over 3,000 books, which they initially dismissed as a scam. This request, sent by an individual claiming to represent '2077AI,' included a spreadsheet of ISBNs for books published between 2020 and 2021 across various academic subjects.

This trend echoes past revelations where AI firms, such as Anthropic, purchased millions of physical books. These books were reportedly subjected to destructive scanning, where bindings were removed for efficient page scanning before the originals were discarded. Legal challenges regarding the use of legally purchased books for AI training have affirmed fair use in some instances.

Companies involved in AI development state their models are trained using a mix of public web content, commercially acquired data, and internal data. They maintain that books are purchased through standard commercial markets and that rare or antiquarian books are not acquired or destroyed. Reports from Germany and Switzerland also indicate similar large, specialized book orders.

Services have been advertised to assist AI laboratories in purchasing large quantities of printed books for language model training. While one company noted such services were exploratory and never launched, the broader search for comprehensive training material continues.