After Anthropic, Amazon also cuts up books to train its AI

Booksellers specializing in old books had been seeing strange orders pile up for months: enormous batches, titles with no apparent logic, and buyers who were consistently anonymous.One of them, in association with American journalists, eventually managed to hide an AirTag inside a book from an order of 1,000 books placed on a marketplace where the buyer can remain anonymous. Weeks later, the tracker came to rest for good in an Amazon warehouse in Las Vegas.

An Apple tracker amid the books, a warehouse at the end of the trail

The area in question goes by the code name VGT3 and even has a logo, featuring a dinosaur clutching a book in its claws. Employees there describe an unremarkable daily routine: cutting off the spines, scanning the pages, and recording barcodes one by one. The books never come back out. The files end up neither as viewable PDFs nor on a storefront: they feed the in-house models. When questioned, Amazon merely says that it “buys books through commercial channels to improve the products and services used by customers”. In loose terms: pass them around. Yet the precedent exists, and it has a code name of its own too. Anthropic had already blazed the trail with “Project Panama,” buying, cutting up and digitizing entire shelves of books to train its models; a federal judge upheld the method in 2025 on the grounds of fair use, notably because the destroyed originals were not resold. Amazon is following the same recipe and scaling it up to industrial levels. The systematic logging of ISBN codes also lends credence to a hypothesis circulating among booksellers: some buyers may be methodically scanning everything that has been printed, number after number.

Why Old Print Is Worth Its Weight in Gold for AI

The web’s reservoir is running dry for those training models. AI giants have already swallowed the bulk of what is floating around online. Recent text poses an unprecedented additional problem: since 2022, a growing share of what is published is itself written by machines. A book printed before that date offers a guarantee nothing else can: human text, never digitized, found nowhere else. Rare books combine all the sought-after qualities—private print runs, uncommon languages, knowledge absent from the web—and their destruction reduces an already thin stock.

Some titles may now survive only as statistical fragments locked inside a proprietary model. Antiquarian booksellers, for their part, are seeing sales climb thanks to these voracious, faceless customers. The method marks a break with the previous era: Google had digitized millions of books for its book-search service without destroying a single one, backed by partner libraries (even if the Mountain View company is not free of allegations on the subject). The AI generation prefers the guillotine, which is faster and cheaper than page-by-page digitization. In Europe, text mining remains governed by a copyright exception coupled with an opt-out right that publishers can invoke; the destruction of rare copies, meanwhile, is nowhere provided for in the law.The reader, for their part, gains nothing: knowledge is leaving public library shelves for private servers.

👉🏻 Follow tech news in real time: add 01net to your feeds on Google and subscribe to our WhatsApp channel.

Source: 404 Media