404 Media Tracked a Rare Book to an Amazon AI Training Facility
A journalist placed a tracking device inside a rare book, watched it change hands, and followed it to an Amazon facility where books are scanned for AI training data and destroyed in the process.
The investigation, published on 17 August 2026 by 404 Media — an independent outlet founded by technology journalists Jason Koebler, Emanuel Maiberg, Samantha Cole, and Joseph Cox — was reported by Emanuel Maiberg. Its title is direct: "We Tracked a Shipment of Rare Books. It Ended at an Amazon AI Training Facility." (404 Media)
The method was physical, not digital. 404 Media suspected that Amazon was buying large quantities of books to feed its artificial intelligence systems. So they placed a tracking device in a rare book they expected Amazon to purchase, then watched the shipment move through the supply chain until it arrived at a facility built for scanning printed volumes into training data. The books do not survive the process. (404 Media)
AI training data is the raw text, images and other material that machine-learning models learn from. The more high-quality text a model ingests, the better it performs at generating language. Printed books — especially older ones — are valuable to AI developers because they predate the internet's flood of AI-generated content, meaning the text is "clean" of what the industry calls AI slop: material produced by other AI systems that can degrade a model's output when fed back in. (404 Media)
The trade in these books involves intermediaries. ISBNdb, a company that sources printed books for AI companies to convert into training data, has reportedly told its clients that "the optics problem is real" — an acknowledgement that buying and destroying books to train AI systems carries reputational risk. (404 Media)
Amazon has not publicly detailed its book-acquisition programme for AI training in the terms 404 Media's investigation describes. The company's name appears throughout the piece because the tracked shipment ended at an Amazon facility identified by the reporters as an AI training operation.
For anyone who has held a rare book — its binding, its age, the particular smell of an old paper stock — the detail that lands hardest is the destruction. The books are not borrowed and returned. They are not digitised and preserved on a shelf. They are consumed. A copy that may have survived decades or longer is taken apart to produce text that a machine will never open again, because a machine does not read the way a person does.
404 Media's reporting connects two worlds that rarely meet on the page: the rare-book trade, with its dealers and collectors and careful provenance, and the data-hungry machinery of large language models, where a book is one more file in a corpus of millions. The tracking device is what bridges them — a small piece of hardware inside a physical object, watching it travel from the world of objects to the world of data.
