Something strange is happening in the world of secondhand books. Over the last several months, independent booksellers have been reporting a pattern of unusual bulk purchases, with scores of novels shipped off to far-away warehouses, and the growing suspicion is that the buyer's appetite is not literary but artificial. A new BBC investigation into the phenomenon traces the boom directly back to the AI industry — and to a court case that made buying books to train models legal under US copyright law.
The BBC report, published August 15 by technology reporter Philippa Wain, lands as AI industry coverage continues to track the expanding hunt for high-quality training data — a hunt that has already moved from web scraping to licensing deals to, apparently, the world's used bookstores.
A week of sales in a single order
In a typical week, Stuart Manley of Barter Books in Northumberland, England — one of the largest secondhand bookshops in Europe — would sell two to three thousand books. Then a single bulk order arrived from a Canadian company that equalled what he would expect to move in seven days.
Manley says he has "never seen the like of this after 30 years in the second-hand book trade." And he is not alone: booksellers from around the world have been reporting similarly unusual mega orders, with no clear indication of where the books ultimately end up. The working theory among sellers is not that distant readers suddenly developed a passion for obscure backlist titles, but that the volumes are being bought for something with a far more voracious appetite for text: AI training pipelines.
Project Panama: "destructively scan all the books in the world"
The trail connecting the secondhand boom to AI firms runs through court documents unsealed last month in the copyright lawsuit three authors brought against Anthropic. Those documents revealed that ingesting old books was referred to internally at the company as "Project Panama" — and that its stated ambition was to "destructively scan all the books in the world."
Destructive scanning is exactly what it sounds like. Books are shipped to facilities where they can be digitized at industrial scale: the spine is removed so every page can be fed through high-speed scanners, and the physical remains are recycled. The book survives as data; the object is destroyed.
The unsealed filings also disclosed that books were being destroyed in the process of training Claude, Anthropic's AI chatbot — a detail that has spread quickly through bookselling communities and fueled the current unease, even among sellers who are profiting from the demand.
David Tobin, who runs Walden Books in north London, has seen the unusual orders firsthand. "In some ways it's very nice to sell some of these titles which haven't been sold for many years, but it would be sad if they are ultimately destroyed," he told the BBC.
The 2025 ruling that opened the floodgates
The legal foundation for the buying frenzy was laid in 2025, when US District Judge William Alsup ruled that training AI software on legally purchased books did not violate US copyright law. The decision came in the lawsuit brought by three authors against Anthropic, and in his ruling Alsup described the company's use of their books as "exceedingly transformative," placing it within fair use under US law.
That ruling — combined with the exhaustion of easily scraped web text — has made physical books one of the most attractive remaining sources of high-quality language data. Unlike much of the modern web, books are edited, coherent and information-dense. For model builders, a pallet of secondhand non-fiction is a clean corpus; for booksellers, it is a sudden, inexplicable gold rush.
Anthropic's response
Anthropic has pushed back on the most alarming readings of the documents. "Claude is trained on a mix of publicly available web data, commercially acquired datasets, and data we generate ourselves," a company spokesperson said, insisting that sourcing books for training is a widely used approach across the AI industry.
"None of our data acquisition programs buy and destroy rare or antiquarian books," the spokesperson added — a carefully worded denial that leaves open the question of common, non-rare volumes, which is precisely what booksellers say is being bought up in bulk.
Whether the mystery buyers are Anthropic's contractors, other AI firms, or data brokers serving several clients remains unknown. Booksellers contacted by the BBC could not confirm the final destination of their shipments, only the pattern: large orders, corporate intermediaries, distant warehouses.
Why it matters
The secondhand book boom is the latest chapter in a broader story about what happens when an industrial-scale technology meets a finite supply of human writing. AI firms have already licensed archives, paid publishers, and litigated the boundaries of fair use. Now they are quietly draining used bookstores — a market built on the assumption that a book's value lies in being read again, not read once, digitized, and pulped.
For copyright watchers, the situation is a strange inversion: the 2025 ruling legitimized buying books for training, but the unsealed documents revealed a destruction pipeline the public never knew existed. For preservationists, the concern is that scarce titles — books with print runs of a few hundred copies — could vanish as objects even as their text lives on inside model weights. And for the booksellers themselves, there is an uncomfortable choice between a windfall and the possibility that the trade's own inventory is being consumed.
"A lot of mystery surrounds Project Panama," Manley told the BBC. Until that mystery is resolved, the world's secondhand bookshops will keep filling orders they cannot fully explain — and watching their shelves empty toward warehouses they will never see.
Stay Ahead of AI Read more AI news →
