Physical bookstores across the United States, the United Kingdom, and Europe are reporting an unusual spike in bulk orders, and the buyers are not readers. They are artificial intelligence companies that acquire physical books to scan for large language model training, then discard the originals.
Barter Books, a used bookstore in Northumberland, England, normally sells between 2,000 and 3,000 books a week. This year, the shop received an order from a company in Canada for that same volume in a single day, according to Fast Company. Other bookshops in the area reported similar demand.
An investigation published by 404 Media traced one such order further. A rare books seller who received a request for nearly 1,000 physical copies allowed reporters to place a tracking device inside one of the books. The package moved from California to Milwaukee, then to Colorado Springs, before arriving at an Amazon-owned facility identified internally as “VGT3.” Workers at the site told reporters their job consists of scanning books all day.
Training AI models on material scraped from the internet has created two problems for developers: legal risk and quality control. Pulling text from pirated online archives, often called shadow libraries, exposed companies to copyright lawsuits from authors, photographers, and news organizations. Buying a physical copy of a book, by contrast, gives the buyer legal rights over that specific item under the First-Sale Doctrine of the Copyright Act, including the right to destroy it after use.
Physical books also solve a data-quality problem. Much of the modern internet now contains AI-generated text, and training new models on that material does not improve them. Books printed before 2022 offer human-written, edited prose with long-form reasoning and narrative structure that short web pages typically lack.
The scale of the practice came to light through documents unsealed in a lawsuit brought by authors against Anthropic. Court filings showed the company had purchased millions of physical books, hired contractors to cut off their bindings, scanned the pages at high speed, and shredded the originals through an internal program called “Project Panama.”
An internal planning document quoted in the filings described the goal plainly: the program was built to “destructively scan all the books in the world.” The same document said the codename existed because the company did not want the effort publicly known.
Court records showed Anthropic had planned to scan between 500,000 and 2 million books under a single six-month vendor contract. Reporting by 404 Media found anonymous bulk orders ranging from 1,000 to 1 million books per purchase.
The destructive scanning itself remains legal. A federal court ruled that training an AI model on copyrighted books is not illegal, provided the company legally acquired the copies it used. A separate judge issued a similar ruling in a case involving Meta.
READ ALSO: Ikole LG Boss Warns Revenue Collectors Against Diverting Council Funds
That ruling has not left Anthropic free of consequence. The company reached a $1.5 billion settlement, one of the largest copyright payouts on record, with more than 300,000 writers who had sued over the practice, according to Fast Company. OpenAI still faces active lawsuits from several parties, including The New York Times, Encyclopedia Britannica, and comedian Sarah Silverman, over similar allegations involving its own training data.
Nobody has an exact count of how many books have been destroyed. Booksellers describe bulk orders for out-of-print and niche non-fiction titles running from hundreds to thousands of copies per transaction, with no indication of a slowdown.
