Amazon Destroying Rare Books to Scan for AI Training Data

When a bookseller grew suspicious about where a 1,000-book bulk order was really headed, 404 Media slipped an Apple AirTag inside the shipment and let it lead the way^[1]. The signal traveled from California through a distribution hub near Kenosha, Wisconsin, an overnight stop in Grand Junction, Colorado, and finally went dark at the north end of an Amazon warehouse complex called LAS8 in Las Vegas^[2]. Inside, a unit Amazon employees call VGT3 - internally represented by a logo of a Tyrannosaurus rex tearing into a book - receives pallets of purchased books, slices off the spines to speed up scanning, and discards the physical copies once the pages have been digitized^[2][3].

Workers on internal forums describe the operation as split across two departments - one that cuts book bindings, another that scans barcodes and pages - with the resulting text reportedly used to train Amazon's Nova AI models^[3]. Confronted with the investigation, Amazon neither confirmed nor denied the AI-training purpose, offering only a boilerplate line that it "purchases books through commercial channels to help develop and improve the products and services our customers use"^[2][4]. The irony has not been lost on observers: a company that built its name as an online bookstore is now running an assembly line that turns physical books into scrap once their text has been extracted^[4].

The choice to destroy books rather than digitize and resell or donate them is not incidental - it is legally load-bearing. In Bartz v. Anthropic, Judge William Alsup ruled that training an AI model on books that were lawfully purchased and then destructively scanned counts as fair use, while training on pirated copies does not^[5]. That distinction effectively rewards buy-and-destroy pipelines: keeping a book intact after scanning, or reselling it, muddies the fair-use argument that the company only extracted a transformative signal from a copy it fully consumed.

Anthropic's own history makes the stakes concrete. The same case that produced the favorable fair-use ruling on destructive scanning also produced a $1.5 billion settlement - the largest copyright payout in U.S. history - over allegations that Anthropic had stored more than seven million pirated book copies^[6]. The lesson other AI companies appear to have absorbed is narrow but powerful: pirating text is a legal minefield, but buying a physical book on the open market and grinding it up after scanning sits on much safer ground^[5].

Underneath the legal calculus is a simpler resource problem: the open web's supply of clean, human-written text is running out. Pre-2022 print books are attractive specifically because they predate the flood of AI-generated content now saturating the internet - one Ahrefs study cited in coverage found that 74.2% of newly created web pages already contained AI-generated text as of April 2025^[7]. A physical book printed before large language models existed is, by definition, human-written and untouched by AI slop, which is why AI companies increasingly value it as a "structurally clean" corpus rather than a novelty acquisition^[7].

That scarcity is tied directly to a technical failure mode researchers call model collapse, in which models trained repeatedly on AI-generated content measurably degrade across generations^[8]. Framed that way, the book-buying operations look less like a publishing curiosity and more like a race to lock down one of the last remaining pools of guaranteed-human text before the rest of the internet is saturated with machine-generated writing^[8].

Amazon's VGT3 operation is not an isolated experiment - it mirrors a pattern already established by Anthropic, whose comparable book-buying and destructive-scanning pipeline has been referenced in coverage as an internal effort in the same vein^[5][7]. Both companies are chasing the same scarce resource under the same legal theory: buy books outright, scan them, and destroy the originals to stay inside the fair-use lane the courts have opened.

Not every major AI lab has followed the same path, at least publicly. xAI has been named alongside Amazon and Anthropic in reporting about companies sourcing old books for training, but the company has publicly stated it does not buy and destroy rare or antiquarian books^[9]. Whether that distinction holds up to the same kind of AirTag-style scrutiny that exposed Amazon's operation remains untested.

Not everyone accepts the framing of this story as cultural vandalism. The buyers reportedly target older, low-value, pre-2022 books sourced through ISBN databases rather than celebrated first editions^[3][7], which raises a fair question about whether "rare books" is even the right frame for most of what's being destroyed. Even so, booksellers like Derek Walker of McNaughtan's Books warn that genuinely rare, long-surviving volumes are sometimes swept into destruction alongside that common stock^[3]. That tension - is "rare" the right word for what's being lost, or is most of it ordinary book-trade attrition wearing a scarier headline - runs through the public reaction to the story as much as the facts of the investigation itself.

Whether the operation is best understood as book burning dressed up in AI branding or as an unremarkable extension of a used-book trade that already discards most of what passes through it, the debate itself says something about how differently people weigh a physical book's value once its text has been copied out of it.