A tracker in a book shipment ended at an Amazon AI training facility

404 Media bought a bulk lot of rare, out-of-print books, hid an AirTag in the shipment, and followed the signal. It ended at a facility Amazon uses to build AI training data. According to the reporting, the books are scanned there and the physical copies are then destroyed.

The part worth understanding is why out-of-print books specifically. Everything easy to scrape has been scraped, so what is scarce now is text no model has ever seen — and a great deal of that exists only on paper, in print runs nobody digitized. The value is the scarcity, which is why bulk lots of old books are worth buying at all. We have not seen a response from Amazon.

The detail worth keeping: Two limits are worth carrying. First, this is reporting, not an admission: what has been shown is where a tracked shipment went and what the reporting says happens there, and destroying books you have bought is not in itself unlawful. The copyright questions around training on scanned text are being fought elsewhere and are not settled by an AirTag. Second, and more useful, is what the story reveals about the shape of the data market. The cheap phase is over. When the remaining supply of unseen text is physical, acquiring it becomes a logistics operation with trucks and facilities rather than a crawler — and that is a market almost nobody outside it can observe. Set that next to a point researchers made to MIT Technology Review this week: the usage and training data these companies publish is the data that flatters them, and there is no independent way to check any of it. The public account of what these systems are trained on is largely the companies' own account, plus whatever a reporter can physically follow.