Amazon Facility Linked to Rare-Book Shipments as AI Training Demand Drives Destructive Scanning

Amazon is using a Las Vegas facility to process printed books for AI training, according to an investigation that tracked a rare book from a bulk shipment to the site. A bookseller placed an AirTag inside one volume, and the device ended at Amazon’s VGT3 warehouse, where workers described receiving large shipments and preparing books for rapid scanning.

The reported workflow removes a book’s binding before imaging its pages, leaving the physical copy destroyed. VGT3 identifies its team with a dinosaur holding or biting a book, an image that has intensified criticism because the symbol echoes the process under investigation. Amazon said it purchases books through commercial channels to improve products and services customers use.

Rare and out-of-print books offer AI developers text that is difficult to obtain online, making them valuable as training data. Large language models have already absorbed huge volumes of web material, while older books provide additional language, subject matter and historical writing that is less accessible through ordinary digital searches.

The timing also matters. Books published before the widespread arrival of generative AI contain text created before models could produce synthetic prose at scale. Training on machine-generated material can contribute to model collapse, a degradation associated with repeatedly recycling artificial data. That pressure helps explain why companies are turning to physical collections, even when scanning requires destroying the originals.

Amazon’s operation fits a wider pattern of AI companies buying physical books, rather than representing an isolated bookselling decision. Booksellers have described mysterious bulk orders for old and uncommon titles, while earlier legal disputes brought attention to similar scanning and destruction practices linked to Anthropic.

Anthropic used purchased books, removed their pages with industrial equipment and scanned them, according to accounts surrounding its $1.5 billion settlement with authors over copyright claims. Those events gave booksellers a reason to suspect that large orders were feeding model-development programs. The Amazon-linked shipment provides a separate, physical trail connecting at least one such purchasing channel to a major technology company.

Destructive scanning turns books with cultural or commercial value into disposable input material. A common paperback can be replaced, but a scarce edition with few surviving copies carries a different cost when its binding is cut and its pages are consumed by an imaging line.

That prospect has alarmed booksellers and collectors, who describe the practice as barbaric and compare it with historical attacks on libraries and written culture. Bulk buying also changes the market before any scanner starts running: sudden demand can raise prices, drain inventories and make it harder for researchers or collectors to locate specific editions. The legal status of buying a physical copy and scanning it remains separate from the ethical question of destroying scarce artifacts, while copyright disputes add another layer of scrutiny.

Amazon’s connection broadens the debate from frontier-model startups to a company with long-standing ties to books, logistics and large-scale fulfillment. Its warehouse infrastructure gives the company a built-in path from commercial purchases to centralized processing, while the VGT3 operation indicates that book scanning can be organized as an industrial task rather than an ad hoc research project.

The investigation also shows how difficult AI supply chains are to see from the outside. A purchase may look like an ordinary wholesale transaction until a tracking device exposes its destination and handling. Amazon’s statement describes the books as inputs for products and services, but the reported destruction of copies places preservation, procurement transparency and data rights alongside model performance in the public discussion.

The immediate questions concern scale, authorization and safeguards for scarce books. Amazon’s public explanation does not disclose how many volumes VGT3 processes, which titles it seeks, whether it preserves digital copies for rights holders, or how it determines that a book can be destroyed after scanning.

Booksellers, collectors and authors now have a clearer target for those questions. Regulators and courts are also examining how copyrighted works enter AI training pipelines, as shown by the Anthropic settlement and continuing industry disputes. If demand for hard-to-find text grows, companies will face pressure to document acquisition practices, protect irreplaceable editions and explain why destructive scanning is necessary.