A Bookseller Hid an AirTag in a Rare Book. It Pinged at Amazon.

TL;DR

A rare bookseller received an anonymous order for 1,000 books on Biblio, got suspicious, and slipped an Apple AirTag between the pages of one volume. 404 Media watched the tracker cross the country and go quiet at VGT3, an Amazon warehouse operation in Las Vegas where, by workers' own accounts, the job is cutting the spines off books, feeding the pages through scanners, and discarding what is left. The anonymous buy-scan-destroy pipeline that has been consuming rare and out-of-print books for AI training finally has a confirmed customer, and it is the company that spent the 1990s calling itself Earth's biggest bookstore.

The sting

Bulk book orders with no name attached have become a pattern in the rare book trade. This time, the bookseller who took the 1,000-book order on Biblio, a marketplace where buyers can stay anonymous, decided to find out where the books actually go. One AirTag between the pages of one volume, and the whole supply chain lit up.

The tracker traveled across the country and stopped at a warehouse in Las Vegas. Not a library. Not a digitization nonprofit. An Amazon facility, where the receiving team's section goes by the designation VGT3 and identifies itself with a logo 404 Media describes as a dinosaur with bared teeth clutching a book. A T-Rex eating a book is, at minimum, honest branding.

Inside the warehouse with the T-Rex logo

404 Media backed the tracker data with worker accounts, and the workers were not coy about what VGT3 does all day.

"I work at VGT3 here in Vegas, and all we do is scan books. Some are assigned to cut books, and others go to receive where they get books and scan the barcodes."

That is the whole operation: receive massive shipments of printed books, log the barcodes, cut off the bindings so the loose pages feed cleanly through high-speed scanners, and convert centuries of print into data. The remnants get discarded. Online forum discussions between Amazon workers, also surfaced in the reporting, match that description.

Amazon did not deny any of it. Its statement says the company "purchases books through commercial channels" to help develop and improve the products and services customers use. Worth noting for the record: the statement never contains the word AI, and 404 Media could not confirm which models, if any, the scans feed. What is confirmed is the operation itself, the anonymity of the purchasing, and the fact that the books do not survive it.

Why the richest companies on Earth are buying paper

The open web has been effectively strip-mined for training text, and an increasing share of what is left is machine-generated. Training new models on the output of old models degrades them, a failure mode documented in Nature as model collapse. Think of it as photocopying a photocopy: every generation gets a little blurrier, and eventually you are training on static.

Rare and out-of-print books are the antidote. They are edited, dense, domain-specific human writing that predates 2022, and much of it exists nowhere online, which makes it invisible to every crawler and every competitor. The bookseller who ran the AirTag sting put it more bluntly: the buyers "just want the content as a bunch of words strung together." The historical value of a book with almost no surviving copies does not appear anywhere in that equation.

Legal, thanks to Anthropic

None of this is a rogue operation, and that is arguably the bigger story. In Bartz v. Anthropic, a federal judge ruled in June 2025 that buying print books, destructively scanning them, and training on the result is fair use, so long as the digital copies are not redistributed. What cost Anthropic its $1.5 billion settlement was piracy, not pulping. The Washington Post detailed Anthropic's own scan-and-destroy pipeline in January.

Then the market did what markets do. In July, 404 Media caught ISBNdb pitching AI labs on bulk print books, 1,000 to 1,000,000 per order, with anonymity and NDAs sold as features. We covered that pipeline when the buyers were still faceless. As of this week, one of them has a face, a warehouse, and a dinosaur logo.

What this means if you build with AI

First, data provenance just got a concrete example with a shipping label. When someone asks where frontier training text comes from now that the web is exhausted, the answer includes anonymous bulk orders on rare book marketplaces, routed to warehouses whose own workers describe the job as cutting up books. If your compliance story leans on "our vendor's data was ethically sourced," this is what sourcing looks like at the margin.

Second, the economics are telling you something. Hyperscalers are paying real money, per physical object, plus warehouse labor, for text that cannot be crawled. Clean pre-AI human writing is now scarce enough to be bought by the pound. That scarcity is the same force behind synthetic-data research budgets and every licensing deal you have seen this year, and it is not going away.

Third, the losses are not abstract. Booksellers told 404 Media back in July that volumes with almost no surviving copies were disappearing into these orders. Under Bartz, that is legal. Whether it should be is a different question, and Amazon of all companies, founded in 1994 to sell books to people who wanted to read them, is now the named test case. Thirty-two years from Earth's biggest bookstore to industrial book pulping is quite the character arc.

Key Takeaways

• 404 Media and a rare bookseller hid an AirTag in a book from an anonymous 1,000-book Biblio order and tracked it to VGT3, an Amazon warehouse operation in Las Vegas.
• Amazon workers describe the job plainly: receive book shipments, scan the barcodes, cut off the bindings, scan the pages. The physical books are destroyed.
• Amazon says it "purchases books through commercial channels" to improve its products and services; the statement never mentions AI, and no specific model has been tied to the scans.
• Rare, out-of-print books are prized training data precisely because they predate 2022 and exist nowhere online, sidestepping AI-contaminated web text and model collapse.
• The pipeline is legal under Bartz v. Anthropic (June 2025): buy, scan, destroy is fair use as long as the copies are not redistributed. Anthropic's $1.5B settlement was for piracy, not pulping.
• In July the anonymous buyers of bulk books were faceless; now one is confirmed to be Amazon, the company that started as an online bookstore.

Sources: 404 Media, TechCrunch, Futurism, AppleInsider