A hidden AirTag slipped into a rare book tracked a 1,000-title order to an Amazon facility in Las Vegas, where bindings are cut and pages scanned before the volumes are destroyed. According to 404 Media, the resulting data is intended exclusively for training artificial intelligence models, in a race where unique content has become a strategic asset.
🔑 Key takeaways
- An AirTag followed a batch of 1,000 books from California to Amazon’s LAS8 facility in Las Vegas.
- According to 404 Media, the books are dismantled, their pages scanned, and the remains discarded.
- Amazon does not confirm AI training and has only issued a statement about its commercial book purchases.
- The sources do not quantify the rise in rare book sales or any reaction in Amazon’s share price.
An AirTag Exposes an Opaque Supply Chain
On August 17, 2026, 404 Media published an investigation establishing that Amazon buys large quantities of printed books, some of them rare, to produce training data for artificial intelligence. For about a year, booksellers had suspected bulk purchases followed by the destruction of volumes after scanning, but could not prove it. A tracker hidden in a book, with the consent of a bookseller, made it possible to document that route and expose a supply chain that had remained opaque.
The tracked book was part of an order of 1,000 volumes. After several weeks, the device led investigators to the LAS8 warehouse in northeast Las Vegas. The route observed was as follows:
| Stage | Location | Observation |
|---|---|---|
| Origin | California | Starting point of the tracked book within the 1,000-volume lot |
| Transit | Milwaukee, Wisconsin | Intermediate stop recorded by the tracker |
| Transit | Grand Junction, Colorado | Additional stop during the tracking period |
| Destination | Las Vegas, Nevada | LAS8 warehouse, northeast of the city |
The route alone does not prove what happens to the contents at the end of the chain. It does, however, establish a material link between a book order and an Amazon site, which separates this investigation from the suspicions previously voiced by the bookselling sector.
A Print-on-Demand Site Repurposed for Scanning
LAS8 normally operates as a print-on-demand book unit. According to 404 Media, however, its northern zone is not dedicated to book production. It is linked to a team called VGT3, which receives, sorts and processes books. The team’s logo shows a dinosaur holding a book in its teeth, which Ars Technica described as a T. rex ready to devour a volume.
Employees interviewed described a precise sequence designed to speed up scanning:
- Receipt of very large shipments of printed books.
- Cutting of bindings to separate the leaves from the block.
- Feeding of pages through scanning equipment.
- Destruction of the book during the process, with its remains sent for disposal.
No incineration is mentioned in the reporting. The destruction described results from dismantling the books, scanning them and disposing of what remains. The method is faster than page-by-page scanning, but it makes every physical copy irretrievable.
Ars Technica also notes that destroying these books to scan them has drawn an increasingly hostile public reaction, particularly from bibliophiles and book professionals.
A Race for Scarce Data in AI
Ars Technica places the operation in a competitive context. Amazon develops what it considers frontier AI models, which would require very large quantities of unique data to stay competitive with Google, OpenAI and Anthropic. These companies generally guard their datasets because they represent a competitive edge. Rare, hard-to-find book texts could offer exactly that advantage.
In her analysis, Amber Neely explains why AI companies are turning to physical sources of text. In her view, three factors converge: the data available online is finite, much of the human content on the internet has already been used to train large language models, and a growing share of new data is produced by AI systems. She also points to the risk of model collapse when systems are repeatedly trained on their own outputs, and to the danger that hallucinations from other models become reference material, with fabricated information later cited as fact.
“Data doesn’t need to be ‘factual’; it needs to be unique and uncontaminated by AI.”
Amber Neely, author of the analysis
Physical books, in her argument, can supply material that has been digitized less often and is therefore more distinctive. The article describes the operation as a “21st-century book burning”, a phrase that reflects the depth of the moral debate surrounding the practice.
Several claims, however, must be separated according to how well they are verified:
| Claim | Source | Status |
|---|---|---|
| Bulk book purchases by Amazon | 404 Media | Documented through tracking of one lot |
| Destruction after scanning | 404 Media, Ars Technica | Described by employees |
| Use exclusively for AI | 404 Media | Reported, not confirmed by Amazon |
| Purchases attributed to Anthropic | AppleInsider | Books concerned not specified |
| Court ruling on destruction | AppleInsider | Jurisdiction and date not specified |
| 10% cost reduction | Forum comment | Not corroborated |
The Anthropic case illustrates this caution. AppleInsider states that the company has previously made large book purchases, yet Ars Technica reports that Anthropic and xAI have publicly said they do not train their models on rare or ancient books. The sources do not specify the titles involved, nor how those public statements fit with any destruction of books. No comparable detail is provided for xAI.
Public Reactions and Open Questions
Amazon declined to comment on 404 Media’s findings. The company did, however, send 404 Media and Ars Technica the same statement, which does not mention AI model training:
“Amazon buys books through commercial channels to help develop and improve the products and services our customers use.”
Amazon statement, shared with 404 Media and Ars Technica
The reactions reported are mainly social and media-driven. Author Sylvia Day shared the Ars Technica headline and link on Threads, a post showing 358 views and no additional comment. On the AppleInsider forum, exchanges between August 17 and 19 focus on the legitimacy of destroying books, the origin of the volumes and copyright law. One contributor argues that current laws force companies to destroy the texts they use, and calls for a rethink of copyright in the AI era.
A comment attributed to Mike Wuerthele claims that removing bindings cuts the cost to 10% and runs about ten times faster than page-turning scanners. This is presented as a personal recollection and is not corroborated by any other source. Another contributor questions whether the books are authentic and warns that inaccurate editions could contaminate training data.
On the financial side, the sources report no movement in Amazon’s share price, no change in market capitalisation, and no transaction, analyst or investor quote. The only market signal mentioned concerns specialist booksellers, who have observed a historically significant rise in rare book sales over the past year, without any figures.
Conclusion: Industrial Opacity Meets the Copyright Debate
The 404 Media investigation establishes one precise fact: a batch of books was shipped to an Amazon site where it was processed for scanning, with the volumes destroyed. The link between that processing and AI model training rests on employee accounts and journalistic reporting, and has not been confirmed by the company, which only addresses its commercial purchases.
Three scenarios emerge. In the first, Amazon clarifies its sourcing framework and processing criteria, which would limit the controversy. In the second, pressure from authors, libraries and booksellers forces legal clarity on destroying originals after scanning, a question on which the court ruling mentioned by AppleInsider remains to be verified. In the third, the practice spreads across other AI players, which would raise the strategic value of rare book holdings and durably reshape that market.
Sources
- 404 Media, investigation into a tracked shipment of rare books
- Ars Technica, Hidden AirTag reveals Amazon is trashing rare books to train AI
- AppleInsider, forum discussion on rare book destruction
- Yahoo News, coverage of the rare book purchases report
- Threads, Sylvia Day post of August 17, 2026
Cet article est publié à titre informatif et éducatif. Il ne constitue en aucun cas un conseil en investissement. Faites vos propres recherches (DYOR) avant toute décision.
