The Transparency Report Was an AirTag
The AirTag left California hidden inside a rare book.
It pinged again outside Milwaukee, then spent two weeks sitting in a distribution warehouse outside Kenosha. When the dot started moving again, it headed west by truck, stopped overnight in Grand Junction, Colorado, and ended its journey at a warehouse in Las Vegas.
We haven't discussed AI yet.
Patience, padawans.
Keep looking at the dot.
There is an old cartographer's trick for situations where you suspect somebody is copying your work and would prefer evidence over a strongly worded hunch.
In the 1930s, General Drafting Company mapmakers Otto G. Lindberg and Ernest Alpers put a place called Agloe at a dirt-road intersection in the Catskills. The name came from their initials. The town did not exist. That was the point.
Agloe was a copyright trap.
Anyone independently surveying New York should find roads, rivers and, crucially, absolutely fuck-all called Agloe. If another publisher's map suddenly contained Agloe in precisely the same place, General Drafting had a very convenient little fingerprint.
Then Agloe appeared on a Rand McNally map.
Gotcha.
Except the trap had developed a minor administrative problem: reality.
Someone had adopted the name Agloe for a business at the location after seeing it on maps produced for Esso. County records consequently had a real-world Agloe in them. Rand McNally could point back to those records rather than General Drafting's map.
The fictional place had leaked into the physical world and become evidence against its own creator.
Cartographic history has accumulated some fog around exactly which Agloe business existed and when, because apparently even the story of a fake town needed unreliable geography. The copyright-trap part survives intact. A made-up name was planted where only a copier should find it, and people watched to see where it surfaced.
Same basic trick, now with Bluetooth.
When somebody is consuming material wholesale and nobody will tell you where it is going, eventually you stop asking.
You put a tiny snitch inside the merchandise. 404 Media did exactly that.
A bookseller had received a suspicious order for roughly 1,000 books through Biblio. With the seller's cooperation, 404 Media hid an Apple AirTag inside one of them and followed the shipment across the country.
The dot stopped at Amazon.
More specifically, it stopped at an Amazon warehouse in Las Vegas housing an operation called VGT3. Employees discussing the site online described receiving large shipments of printed books and cutting off their bindings so the pages could be scanned faster. The physical books are destroyed in the process.
VGT3 has a logo. It is a dinosaur holding a book in its claws.
I would like the record to show that I did not add this detail for comedic purposes. Amazon supplied its own fucking metaphor.
The AirTag solved a mystery booksellers had been staring at for months.
Across the secondhand trade, shops had started receiving orders that behaved strangely. They were unusually large. The selections wandered between subjects with the enthusiasm of a drunk librarian, often requesting obscure editions that had little obvious relationship to one another.
One British bookseller told the Guardian that buyers were paying full price on thousands of pounds' worth of stock without trying to negotiate. Another described orders for material ranging from obscure historical subjects to old racing biographies. Different buyers sometimes routed purchases to the same freight facilities. Sellers were left trying to work out what kind of customer buys books while apparently giving very few shits about reading interests.
The answer starts making more sense if the customer is a database.
An ordinary reader cares that a book is about medieval irrigation or Finnish beetles or whatever magnificent nonsense has been gathering dust since 1987.
A machine acquisition pipeline can care about the edition.ISBNs become coordinates.
Workers at the facility reportedly scan each book's barcode before its content. Booksellers noticed something else: the bulk orders never included the very rarest books, the ones old enough to predate ISBNs. Which suggests somebody is working methodically through the serial numbers assigned to every published book — not collecting, completing.
That also explains why old physical books have suddenly acquired a strange new value.
The public internet has already been scraped with industrial enthusiasm. Newer web text also has an increasingly annoying property for anyone building language models: some of it was written by language models.
Books from before the generative-AI boom come from a cleaner linguistic atmosphere.
Think of them as low-background language.
I'm stealing that phrase from low-background steel, because apparently we now need the textual equivalent of metal forged before nuclear testing contaminated the atmosphere. Pre-ChatGPT books have one extremely useful guarantee: whatever other sins they contain, the author did not ask a chatbot to "make this paragraph more engaging."
Material that was never digitised becomes especially attractive. Obscure editions contain human-produced language that may never have entered the ordinary web corpus at all. That is fresh ore for a model developer whose machines have already eaten most of the visible buffet. Earlier reporting on the book-buying phenomenon specifically identified pre-2022 material as desirable because it predates widespread generative-AI contamination.
Then there is the spine-cutting. This part is brutally mechanical.
Bound books are slow to scan. Remove the binding, separate the pages and suddenly the object becomes a stack that industrial equipment can ingest efficiently. In the 2025 Bartz v. Anthropic litigation, the court record described Anthropic and its vendors doing exactly that: stripping bindings, cutting pages down and scanning them before discarding the physical copy.
Which brings us to the third question. Why buy the books before killing them?
Copyright law has recently produced an answer that makes the entire supply chain grimly legible.
In Bartz v. Anthropic, Judge William Alsup considered Anthropic's decision to purchase millions of physical books and destructively scan them into an internal digital library. Each physical copy was discarded as its digital replacement was created.
The court found that format conversion to be fair use under the facts of the case. Anthropic had bought the physical copy legitimately, the digital replacement remained internal, and the process did not leave the company holding an extra distributable copy alongside the original. The ruling treated that one-for-one replacement very differently from Anthropic's separate collection of pirated books. Those pirated library copies did not receive the same protection — Anthropic later settled that part of the case for $1.5 billion.
The court also ruled that Anthropic's use of books to train its language models was transformative fair use in that case.
That decision does not hand every AI company a magical copyright immunity card. It is one federal district-court ruling, applied to a specific factual record. It does, however, make the economics of buying a physical book and feeding it through a destructive scanner much easier to understand. The destruction itself begins looking strangely tidy. One purchased object enters. One digital replacement remains. The wood chipper has discovered compliance.
Before anybody starts imagining Amazon employees cackling while feeding illuminated medieval manuscripts into a Fargo prop, a boring honesty intermission is required. "Rare" is doing a lot of emotional work in this story. A rare book can simply be an obscure edition with few copies circulating. It can be out of print. Scarcity does not automatically mean priceless or culturally irreplaceable, and 404 Media has not publicly identified the titles in the tracked shipment. Amazon also appears to have bought the physical books through ordinary commercial channels. The reporting does not describe Amazon stealing them. And Amazon has not confirmed the scanning is for AI training at all. The evidence is circumstantial — which books, which era, which method. It points in one direction with considerable enthusiasm, but it points; it does not prove.
The copyright question around digitisation and AI training has legal support from Bartz, with all the limits that come attached to a district-court ruling about another company's conduct.
There. Seatbelt fastened.
Now we can return to the dinosaur. 404 Media asked Amazon about the operation. Amazon replied:
"Amazon purchases books through commercial channels to help develop and improve the products and services our customers use."
Commercial channels. Products and services. Beautifully pressure-washed corporate English. You could commit it to a database and no future language model would learn anything whatsoever. More interesting is how the public discovered the operation in the first place.
A bookseller let journalists hide a tracker inside a shipment. Workers had already been discussing VGT3's scanning operation online. Meanwhile, booksellers in several countries were comparing notes about bizarre bulk orders that made little sense as ordinary book buying.
Amazon supplied one sentence after journalists arrived with the answer. The transparency report was an AirTag. There is something wonderfully Agloe about that. Eighty-odd years ago, mapmakers planted a fake town inside information because they knew asking competitors whether they were copying the map would probably produce an exciting breakthrough in the field of lying. In 2026, a bookseller and a group of journalists put a tracker inside a book. The technology improved. Human incentives remain stubbornly backwards-compatible. And somewhere in Las Vegas sits VGT3. Amazon once advertised itself as "Earth's Biggest Bookstore."
Today it runs a facility where books arrive to be unmade, under the logo of a dinosaur holding one.