Amazon destroys rare books to scan for AI training data

A bookseller who had listed roughly 1,000 books for sale on the marketplace Biblio grew suspicious when a single buyer purchased the entire lot, and agreed to slip an Apple AirTag between the pages of one volume so journalists at 404 Media could follow the shipment ^[1]. The tracker led to a facility inside Amazon's LAS8 warehouse complex in Las Vegas, operating under the internal name VGT3 ^[2]. An employee who works at the site told reporters its sole function is scanning books rather than storing or reselling them: "I work at VGT3 here in Vegas, and all we do is scan books" ^[3]. Workers reportedly cut the spines off books to speed up the scanning process, physically destroying each volume as it is digitized ^[1]. Beyond scanning full text, workers also scan books' ISBN barcodes, and the resulting data is not converted into readable PDFs; it is used exclusively to help train Amazon's AI models ^[4]. Amazon confirmed to reporters that it purchases books through "commercial channels to help develop and improve the products and services our customers use," but declined to confirm or deny that the books are used for AI training specifically ^[5].

The push for physical books is driven by a scarcity problem: as the open internet fills up with AI-generated writing, models trained on that AI-generated text risk "model collapse," a degradation in output quality across successive generations, which is pushing labs toward pre-2022 human-written text instead ^[2]. Printed books are especially prized because much of their content was never digitized in the first place. As 404 Media co-founder Emanuel Maiberg put it, "Printed books are valuable as training data because a lot of the text they contain is not readily available on the internet" ^[6]. This is not a new theory inside the industry - internal documents from Anthropic's earlier book-scanning effort, known as Project Panama, reportedly show a co-founder theorizing that training on books could teach a model "how to write well" ^[6].

The specific act of destroying a book after scanning it traces back to a 2025 ruling in Bartz v. Anthropic, where Judge William Alsup found that destructively scanning a legally purchased print book to build a searchable digital library is fair use, because the resulting digital file simply replaces a print copy the company already owns - a use he called "transformative - spectacularly so" ^[7]. Critically, the same ruling found that Anthropic's separate acquisition of pirated books from shadow libraries was not protected in the same way, meaning the legality of destructive scanning hinges specifically on the book having been purchased legitimately first ^[7]. That distinction still carried a heavy price: Anthropic agreed to pay up to $1.5 billion to settle the underlying authors' class action over its pirated-book sourcing, described as the largest copyright settlement in U.S. history ^[8]. The Authors Guild is among the author advocacy groups whose copyright claims against AI companies over book-based training data resulted in that settlement, which is now a precedent shaping industry practice ^[9].

The story's viral framing - "irreplaceable rare books destroyed for AI" - drew immediate pushback in online discussion. A large contingent of commenters argued that 404 Media never named the specific books in the shipment and cautioned against conflating "rare" with "valuable": much of what moves through bulk used-book channels is unremarkable dead stock that would likely have been pulped or shredded regardless of AI demand. A bookseller participating in the discussion echoed that point, noting that the volume of donated and secondhand books already exceeds what libraries and stores can absorb, so most of it gets shredded either way. A smaller, more vocal group dismissed the framing outright as "ragebait." That disagreement over how much is actually being lost does not resolve the underlying legal question, however: commenters on both sides of the debate pointed back to the Bartz v. Anthropic ruling as the specific mechanism that makes destroying the physical copy - valuable or not - a prerequisite for the digitization to qualify as fair use ^[7].