The Books They Fed the Machine

That is not this story. In July, a bookseller took an order for roughly a thousand volumes through Biblio, a marketplace where the buyer can stay anonymous. At 404 Media's request, the seller slipped an AirTag between the pages of one of them. It surfaced at VGT3, a corner of Amazon's LAS8 facility in northeast Las Vegas, where, according to employees, the work consists of receiving massive shipments of books, cutting the bindings off, and scanning them. Amazon says it "purchases books through commercial channels to improve the products and services customers use." The books weren't bought to be held. They were bought to be input.

The mechanics aren't sinister, just industrial. If you want pages through a scanner at speed, you cut the spine off and sheet-feed the block. It's called destructive scanning because that's an accurate description: cheaper than a cradle scanner, roughly an order of magnitude faster, no trained hands required, and at the end you have a clean digital copy and a stack of loose paper that used to be an object.

Nobody in that building is twirling a mustache. That's the part worth sitting with. This isn't villainy, it's throughput. Somewhere there's a spreadsheet where non-destructive capture failed to clear the bar and the guillotine won on unit economics. It always does. The binding is the friction; remove the friction.

A caution, though, about the word doing the heaviest lifting here: rare. In the trade it generally means out of print or thin on the secondary market — not unique. 404 withheld the titles, so nobody outside the story can say whether these volumes were the last of their kind or the eleven-hundredth surviving copy of a 1974 technical manual. If it's the latter, destroying them removes no ground truth at all, and the alarmed version of this argument evaporates. Amazon plus a guillotine generates dread the facts haven't yet earned, and I'd rather not borrow credibility I don't have.

So here's the version that survives contact with the mechanism: it's about the stock, not the copy. Any single book in that shipment is probably replaceable. But this is a process running at industrial throughput, indefinitely, priced to always prefer the guillotine, drawing down a supply that nothing replenishes. And note what makes a book valuable to a scraper in the first place — out of print, not already online, no digital surrogate anywhere. The selection criterion and the loss criterion are the same criterion. That's not a quirk of this shipment. It's the shape of the operation.

We've watched a gentler version. Google Books, from 2004, went at the same problem and built custom non-destructive rigs; the books went back on the shelves. It's tempting to credit better ethics. It was a property constraint: Google borrowed, so the owner kept a veto, and the owners were libraries holding books in trust for the public. Amazon buys, on the open market, where the commons is not a party and no veto exists.

That's the cut, and it's why the obvious objection is backwards. Amazon has every legal right to destroy books it bought — and that isn't a reason to drop the complaint, it is the complaint. The acquisition channel is the enclosure. Google had to negotiate with trustees and got constrained by them. Purchase routes around trusteeship entirely, and the public with a stake in those volumes has no standing, because standing lives with title. Unit economics explain why the guillotine is cheaper. Ownership explains why nobody can stop it — and only the second tells you where an intervention could ever go: deaccession policy, library sale terms, who is permitted to buy the last physical stock of anything.

Keep the Common a commons. Those inputs were held in common precisely because institutions held them in trust rather than owning them outright, and the transfer from trust to title is where a commons quietly stops existing. Ordinary enclosure privatizes a field; the field is still there, and in principle could be reopened. This privatizes a field and then salts it.

Because what comes out the other end isn't even a corpus. It's weights. A scanned book you can open is a copy — imperfect, but readable, checkable, citable, sitting in a directory with a filename. A book absorbed into a model is a metabolite. You can prompt around it and get back something in its register, but you cannot read it, cannot cite it, cannot audit what the model got wrong about it. And if the physical copy is gone, there's nothing left to check the machine against.

That's the actual loss, and it isn't sentimental. It's epistemic. The book was the ground truth — the thing you could hold up next to the confident paragraph and say: no, that's not what it says.

This won't surface as a scandal. It'll surface as a citation — some model quoting a passage, confidently, from an edition that no longer exists in a form any human can open. Somebody will try to verify it. And the only available verification procedure will be to ask the model again.