# Amazon buys and destroys printed books for AI training data > Source: > Published: 2026-08-17 19:04:35+00:00 Amazon is buying printed books through commercial marketplaces, cutting off their bindings, scanning the pages for AI development and destroying the physical copies, according to an [August 17th investigation by 404 Media](https://www.404media.co/we-tracked-a-shipment-of-rare-books-it-ended-at-an-amazon-ai-training-facility/). [404 Media cofounder Emanuel Maiberg](https://www.404media.co/author/emanuel-maiberg/) identified the operation by following a book across the United States. A bookseller had received an unusually large order for roughly 1,000 volumes through Biblio, a marketplace that withholds buyers' identities from sellers. The bookseller placed an Apple AirTag supplied by 404 Media inside one of the books before shipping the order. The tracked volume left California by air, appeared near Milwaukee Mitchell International Airport and then spent about two weeks at a distribution warehouse outside Kenosha, Wisconsin, the investigation found. It later traveled west by truck, stopped overnight near Grand Junction, Colorado, and arrived at an Amazon warehouse identified as LAS8 in Las Vegas. LAS8 houses several operations. The tracker led Maiberg to the northern section used by VGT3, an Amazon unit represented by a Tyrannosaurus rex holding an open book. Amazon workers discussing VGT3 in employee forums described receiving large book shipments, removing bindings and scanning the pages. Cutting the spine allows loose pages to move quickly through industrial scanners, leaving the printed copy unusable. Amazon acknowledged purchasing books without identifying the products or models that use the resulting data. "Amazon purchases books through commercial channels to help develop and improve the products and services our customers use," an Amazon spokesperson told 404 Media. The statement leaves the destination of the scanned text unspecified. Amazon develops the Nova family of foundation models through AWS and has placed generative AI throughout its retail business, including [Alexa for Shopping](https://www.aboutamazon.com/news/retail/amazon-agentic-ai-gen-ai-shopping). Amazon's [generative AI development disclosure](https://aws.amazon.com/legal/generative-ai-development-disclosure/) describes training sources in broad categories, including licensed, proprietary, public and open-source data, without listing individual datasets or book-acquisition channels. ### Why AI developers want printed books Books contain edited, structured text that may never have been posted online. Older editions also predate the widespread production of machine-generated articles and books, giving model developers a source of human-written material that can be separated from the growing volume of synthetic text on the web. That distinction matters because indiscriminately feeding model-generated material into later training runs can degrade the resulting systems. Researchers have documented this process, called [model collapse](https://www.nature.com/articles/s41586-024-07566-y), in which recursively trained models lose information about the original data distribution. The bulk-buying pattern extends beyond the shipment tracked to Las Vegas. In January 2026, book distributor Ingram told publishers that AI developers were [increasingly purchasing physical books and other media](https://www.publishersweekly.com/pw/print/20260119/99453-ingram-offers-to-not-sell-books-to-ai-companies.html), potentially to scan and ingest them for model training. Ingram offered publishers a way to opt out of known sales to AI developers while warning that it could not always identify the buyer or the intended use. Physical books also carry different contractual constraints from ebooks. Ebook buyers typically receive licenses that restrict copying and redistribution. A buyer generally owns a purchased print copy and can dispose of that object, although creating and using a digital reproduction raises separate copyright questions. ### A court-tested acquisition method Amazon's operation closely resembles a book-digitization program disclosed during copyright litigation against Anthropic. A [June 23rd, 2025 federal court order](https://cases.justia.com/federal/district-courts/california/candce/3%3A2024cv05417/434709/231/0.pdf) documented that Anthropic spent millions of dollars buying millions of print books. Vendors stripped the bindings, cut the pages, scanned them into searchable files and discarded the paper originals. U.S. District Judge William Alsup held that Anthropic's conversion of legally purchased print books into internal digital replacements qualified as fair use on the facts before the court. The order emphasized that each physical copy was replaced by one internal digital copy and that the files were not distributed outside Anthropic. Alsup reached a different conclusion for millions of books Anthropic had downloaded from pirate libraries. The ruling was specific to Anthropic's conduct, but it established a clear incentive for AI developers to buy physical books rather than source unauthorized digital copies. Commercial purchases establish provenance, while destructive scanning produces searchable text without leaving both physical and digital library copies in circulation. Amazon's warehouse adds a significant new participant to that market. The shipment traced by 404 Media shows how a conventional used-book order can enter a private AI data supply chain with the seller unaware of the buyer's identity or purpose. Once scanned, the text can remain useful across future model-development projects. The physical volume cannot return to a bookseller, library or collector. Amazon has not tied VGT3 to a named model or disclosed the number of books processed there. The operation still reveals the underlying strategy: Amazon is converting commercial purchasing power and warehouse labor into proprietary training material that rivals cannot simply scrape from the open web.