There’s a strange, mostly invisible trade happening nearly every day in America: AI companies buying used books by the truckload, slicing off their spines, scanning the pages, and pulping what’s left. Not dozens of books. Not thousands. Orders that run from a thousand to a million books at a time, brokered by intermediaries who advertise anonymity as a feature, because — as one of them admitted in writing — “AI company destroys two million books” is not a headline that generates sympathy.
This should concern Americans. The reason has less to do with books and more to do with a five-hundred-year-old fight over who gets to stand between you and the written word.
First, the mechanics, because they matter. The AI labs have essentially consumed the public internet. What they need now is text that is guaranteed to be written by humans, because training a model on machine-generated text degrades it — researchers call it model collapse, and it works like a photocopy of a photocopy.[1] The internet after late 2022 is saturated with AI output. Ink on paper from before that date is not. Pre-2022 books are the low-background steel of the AI era: the way scientists salvage steel from ships sunk before the first atomic tests because everything forged afterward carries trace contamination,[2] the labs are salvaging human prose from before the first large language models, because everything published afterward might carry theirs.
Physical books have two more advantages. Millions of them were never digitized, so the only copy of what’s in them sits on a shelf. And a 2025 court ruling established that buying a used book and scanning it for yourself is transformative fair use[3] — while pirating the digital version cost one lab a $1.5 billion settlement.[4] So the trucks roll to the used-book warehouses, and the hydraulic cutters run.
The destruction itself is a throughput decision, nothing more. Slicing off a spine turns a book into loose pages that feed through an industrial scanner at speed. Keeping the book intact means turning pages, which means people or robotics, which means cost. At a million books per order, the book stops being a book and becomes an input format.
I run an AI company. I understand the economics of this cold, and I’ll be honest: for a mass-market paperback with ten million siblings in circulation, I don’t mourn the pulping. Used bookstores discard unsold inventory constantly. Not every old book was headed for a loving home.
But these orders are bulk and anonymized, and nobody is checking what’s rare before it goes into the cutter. Booksellers in the U.S. and the Netherlands are reporting out-of-print and genuinely scarce titles moving in these purchases — books that may exist in only a handful of copies anywhere. When one of those goes through the shredder, that’s not recycling. That’s erasure. And the companies doing it possess the most sophisticated cataloging technology in human history. A rarity screen — check the ISBN against worldwide library holdings before destruction, divert anything scarce to preservation — could be built in an afternoon. It hasn’t been. That’s a choice.
Now follow the logic one step further, because this is where it stops being a story about booksellers and becomes a story about power.
For roughly a thousand years before the Protestant Reformation, the Bible in Western Europe existed, for practical purposes, in one authorized form: the Latin Vulgate.[5] Most people couldn’t read Latin. That wasn’t an accident of history — it was an architecture of control. If you wanted to know what the text said, you asked an intermediary, and the intermediary belonged to an institution with its own interests.[6] Men like Wycliffe and Tyndale were persecuted for the crime of translating scripture into the language people actually spoke[7] — Tyndale was executed for it.[8] The fight wasn’t really about theology. It was about disintermediation: whether a person could encounter a text directly, or only through an authorized interpreter.
Gutenberg’s press broke that architecture. Cheap, abundant, physical copies in vernacular languages meant no institution could ever again fully control access to a text.[9] That settlement — the idea that the written record belongs to everyone and survives independent of any single gatekeeper — is so old and so successful that we’ve forgotten it was ever contested.
Here is what worries me: we may be quietly unwinding it.
Consider the end state of the current trade. A rare book — one of the last surviving copies — is bought anonymously, scanned, and destroyed. The physical object is gone. The scan lives in a private corpus that no library, no scholar, no member of the public can inspect. The text hasn’t been published anywhere; it has been ingested. From that point forward, the only way any human being encounters that book is as a statistical influence on a model’s output — paraphrased, synthesized, blended with a trillion other tokens, and served back through a subscription interface.
That is the Vulgate architecture, rebuilt in silicon. A text that exists only as interpreted by an authorized intermediary, behind a paywall, in a form you cannot check against an original — because the original no longer exists. We spent five centuries and a considerable amount of blood establishing that people get to read the primary source. It would be a remarkable thing to trade that away for scanning throughput.
I want to be careful here. Nobody at these labs set out to build a priesthood. This is an emergent property of a thousand individually rational procurement decisions. But architectures of control usually are. The medieval Church didn’t invent Latin to exclude the laity either — the exclusion emerged, and then it was defended, because it was useful.
What makes this frustrating rather than merely tragic is how little it would cost to prevent. Three commitments would do it.
Screen before you shred. Check every ISBN against library holdings; anything below a scarcity threshold gets preserved or donated, not pulped.
Preserve what you scan. If a company destroys the physical copy of a scarce work, the faithful scan — not the model, the actual page images — should be deposited with a library or public archive. The labs lose nothing; the training already happened.
Stop hiding. The anonymity guarantee in this supply chain exists because everyone involved understands the public would object. That’s not a PR problem. That’s a signal.
None of this requires legislation, and frankly I’d rather it never does — Congress writing a book-scanning statute five years from now will get it wrong in ways nobody enjoys. But the industry should understand the stakes correctly. This isn’t about nostalgia for paper. It’s about whether the post-Gutenberg settlement — direct, unmediated, permanent public access to the written record — survives contact with an industry that finds it operationally inconvenient. Buying a book has always meant the right to destroy it[10]; it has never before meant the power to make a text disappear from the human record while keeping its ghost for sale.
The scarcest commodity in AI right now isn’t compute. It’s verified human writing. We are consuming the physical archive of it to build the machines. The bare minimum obligation is that when a book goes into the scanner, something durable and public comes out the other side.
Digitize everything. Destroy nothing irreplaceable. And never, ever make people ask permission to read.
Sultan Meghji is the founder and CEO of Frontier Foundry Corporation and served as the first Chief Innovation Officer of the FDIC. He is a Visiting Fellow at George Mason University’s National Security Institute.
[1] Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal, “AI Models Collapse When Trained on Recursively Generated Data,” Nature 631 (2024): 755–759, https://doi.org/10.1038/s41586-024-07566-y.
[2] The first nuclear detonation, the Trinity test of July 16, 1945, introduced anthropogenic radionuclides into the atmosphere; steel smelted afterward carries trace contamination, which is why shielding for the most sensitive radiation-counting instruments has historically been built from pre-1945 “low-background” steel, frequently salvaged from sunken warships. For an example of a low-background steel counting room in operational use, see M. Manohari, R. Mathiyarasu, V. Rajagopal, V. Meenakshisundaram, and R. Indira, “Calibration of Phoswich-Based Lung Counting System Using Realistic Chest Phantom,” Radiation Protection Dosimetry 144, no. 1–4 (2011): 427–432 (Oxford University Press), https://doi.org/10.1093/rpd/ncq325.
[3] Bartz v. Anthropic PBC, No. C 24-05417 WHA (N.D. Cal. June 23, 2025) (order on fair use) (holding that digitization of lawfully purchased print copies and the use of copies to train large language models were transformative fair uses, while retention of pirated copies was not).
[4] See the settlement in Bartz v. Anthropic PBC, No. C 24-05417 WHA (N.D. Cal.), a class settlement of approximately $1.5 billion covering roughly 500,000 works, announced September 2025 and subsequently approved by the court.
[5] On the Vulgate’s position as the standard scriptural text of the medieval Latin West, see Frans van Liere, An Introduction to the Medieval Bible (Cambridge: Cambridge University Press, 2014), esp. chap. 3; and Richard Marsden and E. Ann Matter, eds., The New Cambridge History of the Bible, Volume 2: From 600 to 1450 (Cambridge: Cambridge University Press, 2012).
[6] On lay religious experience mediated through a Latin liturgy and clerical interpretation in late-medieval England, see Eamon Duffy, The Stripping of the Altars: Traditional Religion in England, c.1400–c.1580 (New Haven: Yale University Press, 1992).
[7] A note of scholarly precision: vernacular scripture was not uniformly prohibited across medieval Europe — translations circulated in French, German, and Italian. The strictest regime was England’s, where Archbishop Arundel’s Constitutions (1407–9), issued in response to Wycliffite translation, made rendering scripture into English without episcopal license a matter for heresy proceedings. See Anne Hudson, The Premature Reformation: Wycliffite Texts and Lollard History (Oxford: Clarendon Press, 1988).
[8] William Tyndale was condemned for heresy and executed at Vilvoorde, near Brussels, in 1536. See David Daniell, William Tyndale: A Biography (New Haven: Yale University Press, 1994).
[9] On printing’s role in ending institutional monopoly over texts, see Elizabeth L. Eisenstein, The Printing Press as an Agent of Change: Communications and Cultural Transformations in Early-Modern Europe, 2 vols. (Cambridge: Cambridge University Press, 1979); and Andrew Pettegree, The Book in the Renaissance (New Haven: Yale University Press, 2010).
[10] 17 U.S.C. § 109(a) (codifying the first-sale doctrine); the doctrine originates in Bobbs-Merrill Co. v. Straus, 210 U.S. 339 (1908).
No posts

Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.