“AI Companies Destroy Rare Books.“ It’s catchy as a headline or a tweet. But it doesn’t survive even minimum scrutiny as a fact claim. Even the literal headlines from trending outlets aren’t great, as written:
"AI firms are buying up old books because they are the last slop-free data left"
"AI Firms Buying and Destroying Thousands of Books To Feed AI Models" (with the lead-in quote “Modern Poisoning Tools“)
The core problems, besides the synthphobic framing? The books in question aren’t “rare” or “antique” (at least in the way described or implied), it’s actually unclear who’s buying all of them, and the processing of the books afterwards actually varies (“destroy” is reductive at best, misleading at worst). Textbook paltering, to put it mildly.
Oh, and that’s before we even engage the relevant counterfactual: what would happen to these books in the event that AI companies didn’t exist? Suffice to say, the facts on the ground don’t point to some idyllic rehoming scenario. In actuality, the very industry that acquires, stores, and sells used books at scale is already in the habit of recycling—and yes, sometimes “destroying”—them on the regular.
But none of that context makes for juicy articles or salacious social media posts. So the truth gets buried, and a more engaging, yet less accurate, narrative gets spun up. Then iterated on, like a synthphobic game of telephone. I’d like to correct the record.
I’m going to work through various claims, but first I need some ground rules and contextual signifers. I’ve codified four main tiers that correspond to the accuracy and strength of a claim. Readers should watch which ones apply where, to see the outrage lives and how much evidence (if anything) there actually is for it.
make the merge itself a point: tier four contains both careful speculation and outright fabrication, and the feed does not distinguish them. That’s the argument, not a housekeeping note.
Tier one — court record. Unsealed filings, including the June 23, 2025 order from the Bartz case. Judges deemed them admissable, then they were adversarially tested; law firms and major press outlets were literally paid to investigate them, some hoping to disprove the claims. The cream survived.
Tier two — business’, government’s, or NGO’s own words. Things like company websites, emails from employees, and internal memos or B2B statements. Authentic as to who said them and why, debatable as straight factual representations of reality.
Tier three — firsthand claims from individuals. Booksellers describing their own order books. Real observation, limited view.
Tier four — everything downstream. Inference, speculation, invention. Book people guessing carefully, journalists extrapolating or posing open questions, and outright confabulations from people with no evidentiary basis attached. These are the least trustworthy, and it’s this tier where most of the claims live.
Rule of thumb for the reader: view anything labeled tier four with caution, and follow my explanations closely; every documented story element is tier one or two, but feel free to fact check for yourself; I am not asserting anything I cannot defend in court, which is a standard some of the people or claims I’ll be referencing…probably can’t meet. When in doubt, check it out. Blind credulity is suboptimal for truth-seeking.
Everything in this section is pegged to tiers one and two. To begin, the AI company Anthrophic maintains a plan or division it calls “Project Panama.” In early 2024, this division circulated an internal doc with language about “destructively scanning” all the books in the world and that they “don’t want it known.” Details of the clandestine operation were revealed in early 2026 when a federal judge unsealed over 4,000 pages of internal company documents and court filings stemming from the copyright class-action lawsuit Bartz v. Anthropic. Project Panama was their response to this case.
That last sentence is crucial, as it goes to an important clarification of the narrative here. While Anthropic drew heavy criticism, and ultimately lost a class action lawsuit, over their previous habit of obtaining free ebooks—including pirated ebooks from websites like Books3, which the judge ruled a gross copyright violation at scale—the company successfully established a novel legal defense for their Project Panama plan under US copyright law. This plan was prompted by complaints about pirating. Whatever one thinks of Antrophic’s overarching goals, there does appear to be a *damned if you do, damned if you don’t* aspect to the dual criticisms here.
Anthropic’s stated reasons for their scanning methods and business plans are a need to get high quality full page scans (which aren’t possible without feeding the pages into the scanner spineless), to scan several books as quickly as possible, and to acquire as many texts as possible to fuel the knowledge base of their Claude AI network. Nowhere does the company state that it intended to deprive anyone else of access to the books or the knowledge within them. It has repeatedly emphasized both Fair Use and First-Sale doctrines to argue that paying for books is legal, as is their choice of what to do with those books after acquiring them. Courts have agreed.
From there, we know details about what Anthropic was beginning to do, but not exact specifics on what actually happened. They intended to bulk buy through commercial secondhand aggregators, liquidation houses, and retailers like Better World Books and World of Books. To maintain secrecy and scale rapidly, the company planed to utilize data brokers and third-party scanning vendors like Datamation Information Services. However, multiple parties (such as The Strand and ISBNdb), either kept silent or flatly denied working with Anthropic, despite being referenced in court documents or news filings. Only Better World Books and World of Books were confirmed.
Books were indeed acquired, as Anthropic admits that pallets of tens of thousands of books were shipped directly to industrial processing facilities, where workers used hydraulic paper cutters to slice off the spines, passed the flat loose pages through production-grade document feeders for optical character recognition (OCR), and sent the physical paper remnants to recycling plants. The unsealed court filings and vendor contracts reveal all this. They do not confirm a single number or tally of how many books Anthropic bought, nor how many unique books were scanned.
This is where we need to announce a tier shift to the realms of tier three (firsthand testimony) and tier four (inference and speculation). All kinds of numbers have been thrown around, and I won’t bullshit you that I have he definitive one. An estimated execution scope of 1.5 to 2 million physical books successfully acquired and scanned into Anthropic’s training dataset as of early 2026 is the likely inference, not a confirmed fact in the public record. That’s worth being blunt about.
While internal planning documents from 2024 framed the ultimate ambition as scanning “all the books in the world” (identifying roughly 40 million titles as readily purchasable), the operational Request for Proposals (RFPs) and contractor ledgers set an actionable baseline target of processing between 500,000 and 2 million titles over the initial operational cycle. The 1.5 to 2 million figure is therefore a reasonable guess.
Yet these figures reflect only what was documented in the judicial discovery record up through the unsealing of the court files in early 2026. Because court discovery records provide a historical snapshot bounded by litigation deadlines, any claims about how many additional physical books Anthropic may have purchased, scanned, or phased out after that cutoff date are completely unverified and would be purely speculative.
Given the available information, we can actually disprove a lot of mistakes and outright lies about Anthropic in particular and AI companies in general. I will handle each of the claims, who made them, and the rebuttals. I will try, as much as possible, to provide direct quotations and to avoid nut-picking randos who aren’t considered serious or important voices in this discussion to begin with. Anyone referenced has a substantial following, is a journalist reporting on the story, and/or is seen as an expert on the relevant domains (books, AI, etc). Fair’s fair, accuracy above all.
The $1.5B settlement was for the buy-scan-destroy book program itself. Quote: "a practice... gained attention after Anthropic paid a $1.5 billion settlement to authors for doing exactly this." Source: David Bentley Hart, relaying an Aaron Parnas report. Facts & Correction: Settlement covered ebook piracy, especially unauthorized downloads from both legal and illegal ebook platforms. Judge Alsup ruled the scan-and-destroy conduct was fair use and imposed no penalty for it.
A judge ruled that destroying the physical book is what makes the practice fair use. Quote: "a federal judge ruled the practice is fair use because eliminating the original means only one copy exists at a time." Source: HedgieMarkets, noted finance Twitter/X account. Facts & Correction: Judge Alsup's holding was that digitizing a lawfully purchased book didn't create a surplus copy. No ruling addressed the destruction itself, since destroying property you own has never required judicial permission. No court docs stipulate as to rarity of books.
AI companies are bulk-buying "rare" books. Source: Numerous, see above. Facts & Correction: The named eyewitness this claim supposedly traces to in the reporting, a bookseller interviewed by 404 Media, has explicitly said the opposite; his bulk orders shared nothing except that every title had an ISBN, and his rare, ISBN-less stock was excluded entirely. Many commentators are linking articles and then giving a direct inversion of their own cited source.
Books are being destroyed "at incredible scale, even if almost no copies remain." Source: Futurism headline. Facts & Correction: No title-level evidence of any specific rare or last-surviving copy destroyed has been documented anywhere in the record, per CryptoSlate's July 28 review of the same underlying sources. Futurism's own body text uses the bookseller's actual word, "uncommon," which the headline discards for "antique" and an unsupported extinction claim.
ISBNdb’s "orders of up to a million books" describes a completed or ongoing deal with Anthropic. Source: IBTimes UK, Dallas Express/Yahoo ("The Vanishing Page"). Facts & Correction: This is misappropriated from boilerplate ISBNdb sales copy about hypothetical or general use business, not a documented transaction with a specific company like Anthropic. No filled order at that scale has been confirmed by any outlet, and as of August 1st, ISBNdb deleted its AI-sourcing pages and stated on its homepage that no such service was ever brought to life.
All of these corrections trace back to tier one and tier two. What we know for certain, again, is that Anthropic began a project to buy and scan books in bulk. It is estimated they acquired and scanned as many as 2 million unique copies by early 2026, but “unique” does not mean “rare,” it means “individual.” Most of the bulk book buys likely contained duplicates, which means they had extras of the same book several times over and needed to specify which was which. Booksellers routinely withhold the actual rare book stock, both for value reasons and because those books lack ISBNs.
To better understand this story, we need a frame of reference for the booksellers themselves. Let’s start with the operations for mass booksellers. Firstly, point of origin—the donation/trade-in bin at physical stores, libraries, and charities. This is how a used book goes from the home of, say, a deceased relative to a Goodwill or used bookstore. This part includes charity shops in the UK/Ireland model, and/or library discard programs and donation-box services in the US. Notably, point of origin does not include warehouses, as that’s the end of the line rather than the beginning.
For now, we’re just starting with collection points. Books arrive as donations or trade-ins, at volumes far exceeding what a single shop can shelve or sell. World of Books alone partners with around 4,000 charity shops across the UK and Ireland, whilst Better World Books runs a discards-and-donations program on US side where it supplies shipping materials and labels to libraries or community drop-boxes. People simply send surplus to be sold, donated, or recycled. Critically, no sorting for rarity happens here. These are volume-in, volume-out points. Workers pull anything obviously damaged; everything else gets boxed or palletized for pickup.
Secondly, bulk purchase and transport to the warehouse. The wholesaler buys the entire undifferentiated stream, often by weight rather than by title. In 2009, for example, World of Books was receiving 140 tonnes of books a week—roughly 300,000 volumes—and bought them at around £75 per tonne. Payment by weight is the key mechanical fact; it means the transaction price is set before anyone knows what’s inside a single box. A first edition and a water-damaged paperback are the same commodity at this stage of the chain. Books travel to the wholesaler’s central facility in bulk containers—large rigid plastic pallet boxes or wrapped cardboard pallets.
Thirdly, is the sorting stage at the warehouse. This is where separation actually happens, albeit in a heavily a automated fashion. At a hub like World of Books’ Coventry facility, every item is fed onto a conveyor barcode-up, scanned, and run through software that checks the item’s circulation levels, then estimates its sell-through rate. Identified items are automatically routed to a swivel-wheel sorter; operatives at outfeed stations drop them into totes based on the software’s sell-through classification, which then get lifted to the correct stock floor.
This is key. Sorting at industrial scale is barcode-driven. A book without a scannable or recognized barcode—which describes most pre-1970 and many rare items—either fails to route correctly or lands in the exception lane: one outfeed station exists specifically to catch unsorted items. Anything else, like books deemed surplus, is sent straight to recycling. That’s right, the booksellers are throwing away or “destroying” thousands of books themselves, at a scale that no AI company could match. But I’m getting ahead of myself, more on those details later.
Simply put, the industrial version of “separating the rare stuff” is really “separating the barcode-recognized stuff from everything else,” and then figuring out what, if anything, in the “everything else” pile is actually worth saving. In most cases, very little actually; that’s why the system defaults toward the recycling stream. Rarity gets protected mostly only when a human intervenes upstream of this, like at a specialist dealer—not at a high-throughput sorting hub.
Fourthly, wholesale distribution. This is where a bulk industrial buyer, like those that Anthropic reportedly worked with, would actually be shopping. Once sorted, the stock is distributed three ways: retail resale (the wholesaler’s own consumer platforms), wholesale to other resellers by pallet or container, or recycling. This wholesale layer is a separately branded market: operators like Zoom Books explicitly sell pre-sorted pallets with documented rejection rates, moving over a million books annually from North American sources into cross-Atlantic wholesale, serving buyers who want volume and predictability. A dealer sourcing this way is buying pallets on spec.
The seller’s own pitch is about consistency and throughput. A supplier who pre-sorts delivers pallets where 85–95% of books are retail-ready, which frames the entire transaction around sell-through percentage, not title-level content. This is almost certainly the layer any AI-adjacent bulk buyer transacts at, because it’s the only layer built for the volumes—hundreds of thousands to low millions—that the parties involved required. Consequently, it’s priced to purpose. The analogy isn’t emptying the shelves of a pharmacy or a museum, but clearing the overflow from a bakery before they end up throwing out a bunch of bread anyway. Important distinction.
Fifthly, the broker, the type of intermediary Anthropic requested to cover their tracks. A broker’s actual function is matching a claim demand—like an AI lab with a target ISBN list and/or a target quantity—to a supply already sitting in wholesale channels like the above. Notably, brokers are NOT sourcing bespoke inventory from individual shops. The broker’s value-add for the buyer is anonymity and aggregation across multiple wholesale suppliers. Nothing in its pitch describes title-level vetting for scarcity or condition. The pitch is about volume, cleanliness of text, and confidentiality.
The unromanticized truth is that bulk book buying and selling is a volume business that includes mass disposal of books, including destruction for pulp via recycling or throwing away copies that end up in landfills. Per World of Books’ own disclosure, they roughly 80% of everything they acquire goes to recycling. That’s over 28,000 tonnes from the roughly 35,000 tonnes of books collected annually, which translates to about 70 million books recycled annually. A measly 1.5 to 2 million books, like Anthropic’s estimated hall, is a rounding error by comparison.
Read that again. It means means four out of five books that pass through this exact company are pulped regardless of what anyone does with the other fifth. And that remaining 20% are still overwhelmingly the book equivalent of day old bread or produce, sold by weight rather than quality or provenance. And that’s before we factor in the books that either these companies, or others, or even private citizens, simply throw away. Roughly 640,000 tons of books end up in a landfill annually in the US. This equates to over 900 million books landfilled per year in the US. (I’ll stipulated that I’m not a math major, so the “number of books” isn’t stated as fact; the tonnage as the solid number and the book-count conversion is illustrative.)
Did I mention publisher-side pulping of unsold new stock? About 2 million unsold new books are pulped every year in a small country like the Netherlands alone. This is a structurally different pipeline—never-sold new stock via the returns system—from the used/donated stock that Anthropic bought from, but it’s useful as a second, independent example of routine, currently occurring destruction at a national scale, with no AI angle at all. The reality is that US-based publishers are likely tossing double, triple, or more of unsold new books, with roughly a third of everything printed by US publishers coming back unsold, and the large majority getting pulped. (Again, millions of books each year, maybe as much as 6 million or more.)
Take any single book, or even a whole title, in this system, exclude out the “AI buyer” story entirely, and what do we end up with? The book’s fate is most likely is unchanged: it sits until it doesn’t sell, then gets pulped, at the same 80% rate that applies to everything else like it at scale. No AI company is required to explain why hundreds of millions of books are destroyed every year. That was already happening, at a scale that makes Anthropic’s 2 million look like infinitesimal. It’s like complaining about a crying baby when a hydrogen bomb is going off at the same time.
Let’s think back to the bread analogy. Day-old bread goes stale whether the bakery starts selling it at a discount or not. A Brioche loaf doesn’t become “rare” just because it’s a unique SKU on the overflow chart that customers pick from. And all this bread was always going to be thrown out if unsold, not preserved somewhere for hypothetical antique bread enthusiasts. Customers shopping at the discount rack cannot, therefore, be described fairly as the cause of the waste. Far from it.
By that same logic, AI companies as buyers aren’t creating a new destruction pipeline. They’re intercepting a fraction of an existing, much larger destruction pipeline one step before the pulper or the landfill. And paying for the privilege, precisely because a court case made buying physical books the most reasonable alternative to pirating ebooks. The marginal effect of an AI company doing a huge bulk book-buy isn’t that good books end up getting destroyed when they otherwise wouldn’t be. It’s that a few books which would be destroyed either way have the knowledge in them preserved digitally, instead of destroyed with nothing preserved at all. A reality which, if anything, inverts the moral framing usually applied to this story.
Perhaps it’s unfair to assume bad faith and accuse those peddling the common claims we discussed earlier of lying. Perhaps the reporters, authors, and other commenters are just lazy, or stupid, or both. But as Dietrich Bonhoeffer argued in his seminal work Letters and Papers from Prison, stupidity isn’t benign. It’s often moral defect rather than an intellectual one, making people easily manipulated and complicit in evil. He wrote that a person becomes “stupid” by surrendering their independent thought to slogans and group pressure. I’ve spoken before about the ills of synthphobia, and I can’t help thinking his words are applicable here. Hanlon’s razor can’t save anyone.
When we outsource our thinking, or engage in telephone games as regards reported facts, we end up poisoning the well. The Internation Business Times talks about AI companies and their endeavors as “Modern Poisoning Tools,“ per someone they quoted. What of those who pollute our epistemic commons? Who tell us things they either know to be untruth or didn’t care to verify before stating as fact, all in service of slandering a company—and, by extension—an entire industry—that they simply have an irrational hate-boner for? I had thought such people already decided that “misinformation” and “disinformation” were high crimes, so what gives?
The reality is that “some used books were recycled after scanning” is a boring supply-chain story. “Rare books were shredded” is a desecration narrative, and desecration is the native register of AI discourse because the underlying objection isn’t to a disposal practice—it’s to AI itself, and to AI companies existing at all. That’s how a verdict drops that incentivizes the very practice now at issue, and how people try to criticize these companies in a heads AI loses, tails synthphobes win fashion. It’s a rigged game.
I don’t entirely blame normies onlookers for being confused tho. The cost of checking is free, but what they might find is still a mess. I wrote this article precisely because a saw a gap in the public record and wanted it filled. If you know anyone who’s laboring under misconceptions in this area, please share it with them. Remember, nine hundred million books a year go into American landfills unscanned, unmourned, and untweeted. Millions more are recycled without preservation of the information. An AI company chooses to do something slightly different and gets pilloried.
I think truth matters more than a spicy or convenient agenda. I hope you do too.
Bartz v. Anthropic PBC, Order on Fair Use (N.D. Cal., June 23, 2025), Judge William Alsup, presiding.
Bartz v. Anthropic Settlement, final approval, U.S. District Court for the Northern District of California (approved July 20, 2026; $1.5 billion settlement covering the LibGen and Pirate Library Mirror shadow-library class, 482,460 works).
Schaffer, Aaron, Will Oremus, and Nitasha Tiku. “Inside an AI Start-Up’s Plan to Scan and Dispose of Millions of Books.” The Washington Post, January 27, 2026. https://www.washingtonpost.com/technology/2026/01/27/anthropic-ai-scan-destroy-books/
Maiberg, Emanuel. “AI Companies Are Buying Tons of Old Books Because They’re Free of AI Slop.” 404 Media, July 21, 2026. https://www.404media.co/ai-companies-are-buying-tons-of-old-books-because-theyre-free-of-ai-slop/
“AI Companies Are Buying Antique Books, Ingesting Their Contents to Train Models, and Then Destroying Them at Incredible Scale, Even If Almost No Copies Remain.” Futurism, July 2026. https://futurism.com/artificial-intelligence/ai-companies-destroying-rare-books
“AI Firms Buying and Destroying Thousands of Books To Feed AI Models.” International Business Times UK, July 2026. https://www.ibtimes.co.uk/ai-companies-demand-printed-books-copyright-concerns-1810831
“The Vanishing Page: AI Firms Scan Then Destroy Rare Book Editions.” Dallas Express / syndicated via Yahoo News, July 2026. https://www.yahoo.com/news/science/articles/vanishing-page-ai-firms-scan-160042724.html
“AI Companies Can Legally Buy and Destroy Books: ISBNdb Reversal Changed Nothing.” Tech Times, August 1, 2026. https://www.techtimes.com/articles/322584/20260801/ai-companies-can-legally-buy-destroy-books-isbndb-reversal-changed-nothing.htm
“AI Firms Are Shredding Physical Books Because Copyright Law Is Quietly Rewarding Them.” CryptoSlate, July 28, 2026. https://cryptoslate.com/why-copyright-law-is-quietly-rewarding-ai-firms-for-shredding-physical-books/
“What the Evidence Actually Shows About AI Companies Destroying Books.” MLQ News, 2026. https://mlq.ai/news/what-the-evidence-actually-shows-about-ai-companies-destroying-books/
“Company That Said It Could Scan and Destroy Books for AI Data-Harvesting Has Deleted That Part of Its Website.” PC Gamer, August 2026. https://www.pcgamer.com/software/ai/company-that-said-it-could-scan-and-destroy-books-for-ai-data-harvesting-has-deleted-that-part-of-its-website-no-such-service-was-ever-brought-to-life/
Scott, Jason (@textfiles.com). “AI Companies Are Scanning and Destroying Old Books.” Boing Boing, July 30, 2026. https://boingboing.net/2026/07/30/ai-companies-are-scanning-and-destroying-old-books.html
@HedgieMarkets. Post on X, July 27, 2026 (3:46 PM), on AI companies and bulk book acquisition.
Hedgie@HedgieMarkets
🦔AI companies are bulk-buying rare books, scanning them through high-speed machines that cut the spines off, and shredding the originals. A service called ISBNdb facilitates orders of up to a million books and keeps buyers anonymous. Pre-2022 books are premium because they're
12:18 AM · Jul 27, 2026 · 23.9M Views
4.11K Replies · 28.7K Reposts · 73.2K Likes
Kehoe, Patrick. Substack Note reposting @HedgieMarkets’ claim, July 2026. https://substack.com/@patrickmkehoe1/note/c-303730870
Bankston, Kevin (@KevinBankston). Post on X, July 28, 2026, 11:59 AM, on the fair-use/piracy distinction in Bartz v. Anthropic.
Carson, Brad. Thread on X, late July 2026, on ISBN-blind bulk orders and the preservation-incentive inversion.
Gioia, Ted. Substack Note, undated, citing an unsourced “400% increase” in rare-book dealer sales.
World of Books Group. Certified B Corporation profile. B Lab Global. https://www.bcorporation.net/en-us/find-a-b-corp/company/world-of-books-group/
“Axiom Automates World of Books Coventry Distribution Hub.” Axiom Group Case Study. https://axiomgb.com/case-study/axiom-automates-world-of-books-coventry-distribution-hub/
“Second Hand Book Wholesalers | UK Sourcing Guide.” Zoom Books, March 7, 2026. https://zoombooks.ca/blogs/second-hand-book-wholesalers-sourcing-guide-uk-retailers/
“World of Books: Revenue, Competitors, Alternatives.” Growjo. https://growjo.com/company/World_of_Books
“Recycle Books Responsibly: Donation, Selling, and Recycling Options.” GreenCitizen, 2026. https://greencitizen.com/blog/recycle-books/
“A Difficult Path: How Volunteers, Waste Workers Save Old Books From Going to the Landfill.” WUFT, January 8, 2024. https://www.wuft.org/business-news/2023-10-31/a-difficult-path-how-volunteers-waste-workers-save-old-books-from-going-to-the-landfill
Simonds, Merilyn. “Where Books Go To Die.” Books UnPacked Blog. https://merilynsimonds.com/books-unpacked-blog/books-go-die/
“Pulping Fiction: Where Books Go to Die.” The Maastricht Diplomat, May 24, 2025. https://www.maastrichtdiplomat.org/post/pulping-fiction-where-books-go-to-die
“Survey on Unsold Books: Returns, Pulping and Recycling.” International Publishers Association SDG Dashboard, August 4, 2024. https://sdg.internationalpublishers.org/sdgaction/survey-on-unsold-books-returns-pulping-and-recycling/
“Impact of Book Publishing on Environment Statistics.” Booketic, February 24, 2026. https://booketic.com/impact-of-book-publishing-on-environment/
Note on sourcing tiers: citations above span the four evidentiary tiers described in the piece — court record (tier one), company/organizational statements (tier two), firsthand bookseller testimony (tier three), and downstream commentary, inference, or speculation (tier four). Readers are encouraged to independently verify any claim, particularly tier-four material, before further circulation.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.