AI companies quietly buying thousands of used books to train models

The email to Dutch bookseller Pieter de Vries offers a concrete example of how AI companies are building physical supply chains to acquire training data. While the public debate around AI training data has focused on web scraping and copyright disputes over digital libraries, this procurement request reveals a parallel, largely invisible operation: the systematic acquisition of physical books through secondhand markets. The request itself—3,001 academic titles shipped to China—bears the hallmarks of deliberate sourcing rather than casual purchasing, yet was designed to look innocuous enough that multiple booksellers dismissed it as spam.

The connection to Anthropic's "Project Panama" provides crucial context. Court records revealed that the company purchased millions of physical books, destructively scanned them, and disposed of the originals—a process that federal courts later sanctioned as fair use. This legal outcome may have emboldened broader procurement efforts across the AI industry. The discovery that ISBNdb, a metadata company, advertised bulk book sourcing services for "LLM training needs" at scales of "1,000 to 1 million books" suggests that infrastructure for large-scale acquisition is being built, even if ISBNdb later claimed the service was never operational. Similar orders reported in Germany and Switzerland indicate this is not an isolated incident but part of a coordinated, international supply strategy.