An AirTag in a book shipment reveals what legally clean training data actually costs

An Apple AirTag hidden inside a shipment of rare books led an investigation to Amazon's VGT3 team at a warehouse in Las Vegas, where employees cut the bindings off printed books so the pages feed a scanner faster, destroying the book ^[1][2]. The investigation, published Monday by 404 Media, is the first to publicly pinpoint where bulk book purchases tied to AI training physically end up ^[3], which turns a year of bookseller suspicion into an auditable supply chain.

Amazon said it "purchases books through commercial channels to help develop and improve the products and services our customers use," and said the operation had not been reported before ^[4][5]. The team's logo is a dinosaur clutching a book in its claws ^[6].

The scale is the story. Anthropic ran a comparable programme, Project Panama, described in internal documents as an "effort to destructively scan all the books in the world," digitising millions of books ^[7][8]. One vendor's project proposal requested 500,000 to two million books over six months ^[9]. That is roughly 2,800 to 11,100 books a day, every day, from a single request ^[10]. For comparison, the source notes that the upper end alone would cover the entire physical catalogue of Montreal's Grande Bibliotheque, estimated at 1.2 million books ^[11].

Why paper, and why now. Books printed before 2022 contain text that is not mostly available online and, critically, was not written by a machine; training on machine output risks model collapse, where quality degrades with each recursive loop ^[12]. Large AI companies are running out of useful text ^[13]. A cottage industry has formed to supply labs with physical books that get stripped, scanned and discarded ^[14].

The economics only make sense alongside the case law. A federal judge ruled that training on legally purchased books can qualify as fair use, a partial win Anthropic, Meta and OpenAI all claimed ^[15]. Pirated copies were treated differently: Anthropic agreed to a $1.5 billion settlement over scanned pirated titles, and Salesforce now faces a class action over alleged book piracy ^[16][17]. Buying and shredding a physical copy is therefore not sentiment or sloppiness. It is the cheapest available proof of provenance, and the destruction is what you pay for a receipt.

The market has already priced it. Booksellers noticed a spike in bulk orders over the past year from buyers who were not price-sensitive and picked titles seemingly at random ^[18], and secondhand sales are reportedly booming, which some owners tie to AI bulk buying ^[19]. One bookseller told the BBC the world "no longer needs five million copies of The Da Vinci Code" ^[20]. The same appetite is showing up outside books: Google and Mercor fought over access to Spirit Airlines' old data, including hundreds of millions of internal emails, Microsoft Teams chats and employee productivity records, and Google won ^[21].

What to watch. First, whether other labs' scanning facilities get named the same way, now that a $30 tracker has proved the method works ^[1][3]. Second, whether vendors of the 500,000-to-two-million-book variety start pricing on scarcity rather than volume, since indiscriminate title selection suggests buyers have not yet hit a supply constraint ^[9][18]. Third, the Salesforce class action, which will test whether the purchased-versus-pirated line drawn so far holds as a durable liability boundary rather than a one-off settlement number ^[15][17]. For anyone selling data, the operative fact is that clean provenance, not volume, is what commands a premium.