Secondhand book sales are soaring, with many independent booksellers observing unusual bulk orders being shipped to warehouses, primarily from AI companies. Stuart Manley of Barter Books, who typically sells 2,000 to 3,000 books weekly, recently received a single bulk order from a Canadian company for a quantity equivalent to his usual weekly sales. David Tobin of Walden Books in North London has also seen unusual sales patterns, indicating a shift in the market.
This boom is largely attributed to a 2025 US court ruling that found using purchased books to train AI software did not violate copyright law, deeming such use "exceedingly transformative." While beneficial for clearing old inventory, booksellers like Manley and Tobin express discomfort with the potential destruction of these books. Concerns escalated after court documents revealed that books were being destroyed during Anthropic's AI training process.
Destructive scanning involves removing a book's spine to rapidly scan its pages, with the remains often recycled. While some booksellers, like Derek Walker of McNaughtan's, believe the loss of certain academic texts with many existing copies might not be significant, others worry about irreplaceable works. Walker notes they have sold books that are the only known surviving examples of 18th-century editions. The practice creates an ethical dilemma for booksellers, balancing increased trade with the potential loss of historical or intellectual value, especially for rare or out-of-print titles. ISBNdb, a service facilitating bulk purchases for AI companies, acknowledges the "optics problem" of AI firms destroying millions of books.
A recent investigation by 404 Media, using an AirTag, tracked a rare book from a bulk order to an Amazon AI training facility in Las Vegas. This facility, whose logo depicted a T-Rex devouring a book, housed a team dedicated to tearing spines and scanning pages. This confirmed suspicions that major tech companies like Amazon are involved in destructive scanning to acquire the vast amounts of text needed for large language models, particularly for rare or untranslated books that offer unique training data not readily available online.
