Secondhand booksellers across the UK, Ireland, and North America are reporting an unprecedented wave of erratic, bulk orders from opaque buyers, sparking widespread speculation that artificial intelligence developers are acquiring physical volumes to harvest their text for machine learning.
Independent shop owners described receiving requests for unusual combinations of titles that defy typical purchasing patterns. Stuart Manley, co-owner of Barter Books in Northumberland, noted that recent orders included an Estonian translation of a John le Carré novel, a specific imprint of Anne Brontë’s Agnes Grey, and an issue of a 1980s naval magazine. Sellers in Ireland and Bath reported similar surges of sporadic, machine-driven purchases covering subjects as disparate as 18th-century African farm tools and mid-century racing drivers.
Rather than seeking standard trade discounts, the mystery buyers are paying full price and requesting bulk shipments delivered to central logistics hubs, such as freight warehouses near Heathrow Airport. The opaque nature of the transactions has disrupted traditional trade dynamics, with experienced booksellers noting that multiple pseudonymous accounts often route shipments to identical addresses.
The flurry of activity points to tech companies' growing appetite for pristine human-written content. As AI developers exhaust digital data sources, pre-2022 physical books have emerged as prized assets because their text remains entirely unpolluted by AI-generated content. Acquiring physical books allows firms to digitize niche or out-of-print literature that was never published online.
Industry reports have previously revealed that leading AI firms, including Anthropic, spent tens of millions of dollars acquiring physical books, unbinding them to scan their contents before sending the paper for recycling. While Anthropic stated that its programs do not destroy rare or antiquarian titles, forum discussions among booksellers express growing dismay that thousands of ordinary physical books are being purchased merely to be scanned and pulped for training data.