What the Anti-AI Viral Story Gets Wrong About Amazon’s Destructive Book Scanning

A viral LinkedIn graphic turned a good investigation into a cleaner and angrier story. In that version, Amazon secretly bought thousands of precious rare books, carried them into a high-security AI laboratory, shredded them, and disposed of the evidence. The picture supplied antique volumes and a laughing corporate villain. The reporting supplied neither.

The underlying event is real. Emanuel Maiberg and 404 Media tracked one book with an AirTag from a bookseller’s bulk order to Amazon’s LAS8 complex in North Las Vegas. Worker accounts described a nearby operation called VGT3, where teams received books, removed their bindings, scanned the pages, and discarded the physical remains. Maiberg later said Amazon confirmed that AI training was one purpose. That is a serious finding. It does not need the additions.

My judgment is narrow. The 404 Media investigation is credible and appropriately qualified. The viral presentation is anti-AI propaganda because it changes the reported quantities, invites a false picture of the books, adds an unsupported concealment motive, and presents inference as direct observation. Correcting those claims does not excuse Amazon. It leaves the actual preservation and transparency questions easier to see.

Scope note: I am comparing the original investigation, later reporting, and the public claims about one tracked shipment and Amazon’s VGT3 operation. The available evidence does not establish whether every scan enters an AI dataset, whether every later use is lawful, or how many scarce physical copies Amazon has destroyed.

What the AirTag established

The tracked order contained about 1,000 books, not several thousand, and the anonymous bookseller did not devise the tracking experiment alone: 404 Media proposed it, supplied the AirTag, and asked the seller to conceal it inside one book. The tag traveled from the seller through a Wisconsin logistics warehouse to the north end of Amazon’s LAS8 complex in Nevada, while GeekWire’s separate request for comment produced the same broad response that Amazon had given 404 Media.

This route is strong evidence that the tagged book reached Amazon’s complex, but the tag could not report what happened after the package entered the building or establish the fate of every other book in the order; the connection to VGT3 came from the tag’s north-end location and posts attributed to Amazon workers who described VGT3 as a separate operation there.

That method is considerably better than a bookseller guessing the identity of an anonymous buyer. It is still a chain made from tracking data, location, worker posts, and Amazon’s response rather than direct access to the production floor.

The later employee account strengthens the scanning claim

Nine days after the first story, 404 Media published an interview with an anonymous worker who had been inside VGT3. The worker described receiving stations, barcode checks for duplicate books, machines that cut off bindings, high-speed page scanners, and large open containers holding loose pages after scanning. The account matches the earlier worker-forum descriptions and gives the destructive process more support than the first article had on its own.

The interview remains one anonymous source, edited by the reporter, without photographs or outside inspection of the production line. Readers can reasonably give it weight without treating it as a complete operational record. Amazon has not published the process, the number of books handled, the title-selection rules, the retention policy, or the destination of the scans.

“Thousands,” “rare,” and “shredded” change the story

The viral account makes three substitutions that appear small until their effects collect.

First, it changes an order of about 1,000 books into “thousands.” VGT3 may handle thousands of books across many shipments; the later employee account describes a substantial operation. The tracked seller’s order, however, was not several thousand, and the reporting did not establish the total number of books consolidated with it before delivery.

Second, “rare” does not mean what the antique-book montage asks the viewer to imagine. In a later public-radio interview, Maiberg described the reported orders as involving obscure, low-circulation, vanity-published, and obsolete instructional material rather than famous first editions. The later employee interview separately described books in several languages. These books may be scarce because demand was low, not because collectors consider them treasures. The seller’s title list remains confidential, so nobody outside the investigation can check the condition, value, or surviving copy count of each book.

Third, the evidence supports destructive scanning, not the vivid claim that complete books were fed into a shredder. Workers described cutting off bindings, scanning loose pages, and discarding the pages into bulk containers. The distinction matters because a person reading “shredded rare books” beside a photograph of antique bindings will form a specific scene that the sources do not establish.

The secrecy and evidence claims were added later

LAS8 is a known Amazon print-on-demand warehouse, and VGT3 appears to be an internal operation within the same complex. Amazon did not announce its purpose publicly, but no report establishes that VGT3 was an unusually high-security laboratory or that its identifier was a covert project codename. An internal warehouse identifier does not establish a classified program.

The claim that Amazon destroyed books to “dispose of the evidence” is more serious and has no reported basis. It assigns a concealment motive without a source. Destructive scanning has an ordinary operational advantage: a worker can feed loose pages through an automated scanner much faster than a bound volume. A possible legal advantage also exists, but it requires careful wording.

The 2025 Bartz v. Anthropic fair-use order considered Anthropic’s conversion of lawfully purchased print books into digital library copies. Judge William Alsup held that the one-for-one conversion was fair use on the facts before the court. Anthropic bought books, removed the bindings, scanned them, discarded the originals, and retained digital replacements for storage, search, and research. The same order rejected Anthropic’s fair-use defense for books acquired from pirate sites. The later $1.5 billion settlement received final approval in July 2026 and concerned the piracy claims, not a universal license to scan any purchased book for any purpose.

The opinion gives a company one reason to think that destroying the physical copy may support a one-copy-for-one-copy legal argument, but it does not make destruction mandatory, validate every downstream use, prove Amazon’s motive, or establish that Amazon adopted Anthropic’s process or reasoning.

AI training is well supported but not publicly explained

The evidence for an AI-training purpose is stronger than rumor and weaker than a detailed public admission. Maiberg has said that Amazon confirmed AI training was at least one use of VGT3. Amazon’s attributable statement says only that it purchases books through commercial channels to help develop and improve customer products and services. GeekWire asked directly and reported that Amazon would not publicly identify AI training as the reason.

Independent evidence supports the broader practice: Inc. found Amazon Seller Forum discussions dating to 2024 in which booksellers reported unusual bursts of orders routed to LAS8, including 68 orders for one seller and 48 overnight university-press orders for another, and an Amazon forum representative said the orders were valid without connecting them to VGT3. Ingram Content Group now gives publishers an AI-training opt-out for sales it knowingly makes directly to AI companies. Those records show that physical-book acquisition for model development is an industry practice. They do not disclose what Amazon does with each scan.

Destructive scanning creates specific preservation losses

Digitizing a book preserves a representation of its printed text. It does not preserve the complete physical object. A binding, inscription, set of marginal notes, library marks, ownership history, inserted document, or unusual production detail may disappear when workers cut the book apart and mix its pages with pages from other destroyed books. If only a few accessible copies remain, destroying one can also reduce the chance that a future reader or researcher can inspect it.

That concern depends on facts not publicly known: the number of surviving copies, the condition of the purchased copy, whether a trusted archive already holds one, the quality and retention of the scan, and whether Amazon screens material before destruction. The confidential seller list prevents outsiders from checking the tracked titles, while Amazon has not disclosed its broader title-selection or preservation rules. “Amazon is destroying human knowledge” avoids all of those questions by treating the printed object and its textual content as identical, then pretending both vanish. The actual loss is more specific and, for some books, could be substantial.

This is why the propaganda is not merely embarrassing decoration. It makes a responsible critic defend claims the investigation never made, while Amazon can answer the exaggeration instead of the evidence. The better demand is plain: disclose the program’s scale, selection rules, preservation checks, scan destinations, and uses. Until Amazon answers, the public record supports criticism of a large destructive-digitization operation with uncertain preservation controls, not a story about priceless books entering a covert shredder so the company can erase the proof.