The practice of guillotining books - that is severing the binding from the printed pages such that the latter can be scanned, shredded, and pulped - originally came to light thanks to the Bartz v. Anthropic ruling last year.
The ruling was a mixed bag. It led to multibillion dollar settlements, yielding four- and five-figure paydays for authors whose pirated works were scraped off the internet and dumped into the corpuses of commercial LLMs. More consequentially, it established a precedent for regarding this practice as a violation of copyright and therefore forced the frontier model developers to start negotiating with publishers.
This precedent has not been universally applauded by professional writers. As Cory Doctorow argues in The Reverse Centaur’s Guide To Life After AI (2026), it essentially created a framework for developers to strike deals with publishers, but created no parallel inducement for workers - writers, editors, illustrators, designers, etc. - to have a seat at the negotiating table. It turned media conglomerates and tech giants into a powerful alliance in the extraction of creative labor and the busting of unions in the creative industries.
It was as part of the discovery process for Bartz v. Anthropic that Project Panama became part of the public record. Project Panama was, as Anthropic itself put it in internal documents, “an effort to destructively scan all the books in the world.”
Anthropic realized this sounded like something The Joker would say. So they instructed their employees: “You should avoid talking about it in public areas, and the fact that we are working on this should not be shared with anyone outside Anthropic.” The industrialized guillotining of books during a worldwide resurgence of fascism was quite taboo, they realized, but it was not illegal. Right?
In Bartz v. Anthropic, Judge William Alsup decided that it was not. Alsup cited a well-established exception to copyright for cases of “transformation” by the owner. As a podcaster who sometimes creates audio montages I am familiar with this exception and grateful for its reaffirming. But Judge Alsup was also, knowingly or otherwise, sentencing millions of books to the guillotine.
Alsop’s argument, in brief, is as follows. Anthropic obtained the print books legally (either by purchase or remaindering from libraries), transformed them into digital tokens which are transmitted in a different media, and in the process “the print original was destroyed.” The latter stage of the transformation is essential, according to Alsup, because it prevents Anthropic from expanding the quantity of salable copies on the market and thus cutting into the revenues of the copyright holders.
It is not my intention to critique Alsup’s judgment, though some legal scholars and political economists have. Regardless of the quality of Alsup’s reasoning, his judgment has since created a perverse incentive. He signaled to developers that it is illegal (and therefore costly) to freely obtain and transform digital texts, which, so long as they remain freely accessible, are almost infinitely reproducible, while simultaneously signaling that it is legal to obtain, transform, and destroy print texts, which are, in many cases, finite.
By pirating digital books Anthropic (and many of its competitors, at least as alleged in both settled and ongoing litigation) has not changed the fundamental “stock” of digital books. By guillotining books it had. And in the case of books which were rare, out-of-print, undigitized, and/or published by defunct presses, it had put the work a step closer to extinction.
The frontier developers understood Alsup’s decision perfectly well. He had, in a fashion characteristic of neoliberal governance, turned the legal limbo of corpus scaling into an arbitrage calculation, one which could be computationally modeled. In the procurement of text, developers had three options: 1.) scrape and settle, at the estimated cost of $3,500 a book plus legal fees, 2.) negotiate purchase of corpus development rights (potentially with exclusivity clauses) directly from publishers, or 3.) legally obtain and guillotine print books.
Developers set about doing all three, likely choosing the procurement strategy for each publisher or even individual book based on proprietary financial models. For at least some of the frontier model developers, this meant ramping up their book extradition and execution operations, while also accounting for the fact that Anthropic had been proven correct. Planning to “destructively scan all the books in the world” is “not a headline that generates sympathy,” as ISBNdb put it as part of sales pitch to keep AI companies anonymous by subcontracting procurement. ISBNdb quickly discovered how right they were, deleting the pitch from their website and issuing a public apology.
Several news organizations, notably 404 Media, have been tracking the book extraditions and executions to remote and nondescript warehouses, like the VGT3 facility in exurban Nevada, owned by Amazon, which features this rage-inducing logo:
I can only interpret the dinosaur as taking pleasure in the imminent extinction of the book, which mirrors his own.
In some cases, booksellers have been doing their own sleuthing, attaching unobtrusive tracking devices to suspicious orders from anonymous customers, hoping to prevent themselves from unknowingly sending large numbers of irreplaceable books to be guillotined. One such bookseller, who spoke with the Wall Street Journal, tracked shipments to both Amazon and Meta facilities, and from there one surreptitious tracker made it, presumably still attached to a decapitated spine, all the way to a paper recycling facility in Mexico.
While acknowledging the outrage people feel at these operations, an outrage which drives traffic to news sites, the Wall Street Journal reporting has been characteristically unwilling to name the actual harm book-guillotining is doing, instead insinuating that it is a victimless practice that merely offends soft-hearted people’s sentiments. WSJ reporters Mellissa Korn and Jessica Mendoza both describe physical books as possessing a kind of “magic” which stirs nostalgic fondness because of “that feeling you get when you open a book for the first time” or the “wonder” of “who owned this before and what’s their story.” But, Korn clarifies, “Is this going to completely destroy the used book industry? No.” In fact, “rare book sales are booming” precisely because “they’re getting sliced up and fed to AI.”
And to some extent I agree with them. We are a long way from Bradburyian dystopia. There are billions of volumes of books in circulation, with millions of new titles published every year. It is undeniably a technofeudal fantasy of frontier developers like Sam Altman and Elon Musk to enclose the general intelligence of humanity in a statistical model, accessible only by chatbot, and censorable by them personally. But that is, at least for now, still a very far-fetched fantasy.
However, the books that developers are most eager to extradite, execute, and enclose the contents of are the ones at greatest risk of extinction. The WSJ reporters acknowledge that there are an increasing number of titles on sites like Biblio and Abebooks for which there are fewer volumes for sale than there are frontier AI developers. The cost of those titles is sky-rocketing, as anybody who has been regularly purchasing used books online can testify. Much as data centers are pushing up the price of electricity for surrounding communities, corpus scaling is pushing up the price of out-of-print books for researchers.
The contraction of used book stock and the inflation of used book prices is happening simultaneous with the reduction of library collections, the defunding of physical archives, and the enshittification of digital archives, and all of the above are at least partially caused by the same pressures from frontier developers. It is a cruel detail of Anthropic’s Project Panama that it was overseen by an executive who previously worked for Google Books, a once-vital digitization project that is increasingly non-functional.
It is humanities scholarship that will suffer most immediately, as it is more likely to depend upon obscure texts, specific editions, and material books, and it relies on methods of quotation and citation for which LLMs offer no alternative. The cost of the Alsop decision has already been millions of books. It is highly likely that some idiosyncratic text has already been lost to the world for good, now nothing more than tokens on the silicon of post-history.



