AI companies are quietly buying up millions of physical books, including rare and obscure titles, slicing them apart to scan them for training data and destroying them, under a secret programme deliberately kept from the public
By The Conspiratory EditorsAugust 4, 2026
Where the evidence lands: ProvenThat AI companies have been secretly buying millions of physical books, including rare and obscure titles, destroying them by slicing them apart to scan their contents for model training data, and deliberately concealing the programme from the public.
Proven. The central claim is documented, and the documentation is the company's own. Internal records unsealed by a judge in a copyright lawsuit and made public in January 2026 show that Anthropic ran a confidential programme it called Project Panama, described in an April 2024 memo as its effort to destructively scan all the books in the world. The same memo explains the codename: the company did not want it known that it was working on this, and told employees not to discuss it outside the company. Books were bought in bulk, sliced at the spine so pages could be fed through a scanner, and not reassembled. An internal memo from October 2024 indicates millions of books had already been purchased. In an August 2024 exchange with a bulk bookseller, the executive leading the project agreed that less common books were a good place to start. Two parts of the viral version are not established and this file does not assert them: whether other AI companies ran comparable programmes could not be confirmed, and what less common meant in practice, as against genuinely rare or collectible titles, is unclear. One point is frequently missed and belongs in any honest account: a federal judge ruled that destructively scanning lawfully purchased books was fair use, so this was found to be legal. The rating here is that the described conduct occurred, not that any law was broken.
Believed by: A wide audience with no single political character: writers and publishers, librarians and rare-book communities, AI critics, and a large general readership who encountered it as an emotionally striking story about books being pulped
The full story
The sentence that did the damage
Most viral claims are built on an inference. This one is built on a sentence somebody at the company actually wrote.
An internal memo dated 13 April 2024 opens by defining the work: Project Panama is our effort to destructively scan all the books in the world. A few lines down it explains the codename: we use a soft codename for it because we don't want it to be known that we are working on this. Employees are told the document is internally visible but should not be discussed in public areas, and that the existence of the work should not be shared outside the company.
The memo is not a leak. It became public because a judge in a copyright lawsuit brought by authors ordered the company's internal records unsealed, and it has been readable since January 2026.
That is why this file carries the rating it does. The core of the viral claim, that an AI company bought physical books in bulk, sliced them apart to scan them, did not put them back together, and deliberately kept the programme quiet, is documented in the company's own paperwork. What follows is about the parts the posts added, and about one crucial part they usually leave out.
What the file actually contains
The unsealed material is more specific than most reporting conveys.
The programme was run by Tom Turvey, who had previously worked at Google on partnerships including Google Books, which is a relevant piece of professional history: mass book digitisation is a recognised discipline with two decades behind it.
An email exchange starting 27 August 2024 with Wonder Book, a bulk bookseller, has Turvey confirming that less common books are a great place to start. Other correspondence shows approaches to booksellers and publishers in other languages, including the Spanish publisher Oceano.
An October 2024 internal memo indicates millions of books had already been bought by that point. The precise number is redacted, but the M at the end of the redaction survives and fixes the order of magnitude. A separate 211-page partially redacted document lays out the process, including how titles were categorised.
The physical mechanics are worth stating plainly because they are the heart of why people reacted. To scan at that rate, a book is cut at the spine and the loose pages are fed through a sheet-feed scanner. Nothing puts the book back together afterwards.
Where the posts went past the record
Two additions travelled with the true part, and both matter.
The first is the plural. The posts said AI companies. The documents concern one. Snopes rated the rumour Mostly True rather than True for exactly this reason: it could not confirm that other AI companies had run comparable programmes. The 404 Media investigation that set the wave going described a broker marketing bulk physical-book acquisition to AI developers, which shows there was demand in the industry. It does not show that any named competitor was slicing spines.
The second is rarity. “Less common” is the phrase in the email, and it is doing a great deal of work in the viral retelling, where it becomes rare, antique, irreplaceable. In a data-acquisition context it may simply mean poorly represented in existing digital corpora, which is a property of a title rather than of a physical object. The one piece of evidence pointing the other way is ambiguous: a process document uses a Book Industry Study Group category called Antiques and Collectibles as a categorisation example, and nobody can currently say whether that means antique books or books about antiques. Those are very different things. Snopes asked the company and reported no answer.
Nobody has shown that a first edition was pulped. That does not make the programme benign; it makes the specific accusation unproven, which is a different thing.
The ISBNdb thread deserves the same care. A now-deleted page on the company's blog described the practice and contained the line about two million books returned to pulp, alongside a strikingly candid observation that such a headline does not generate sympathy. The company later said it has never purchased, scanned or sold a book for AI training or anything else, that it does not train AI models, and that the page was a test of market interest for a service never launched. Both can be true. A deleted marketing page proves somebody saw a market. It does not prove they served one.
The part that keeps getting dropped
One fact is missing from almost every viral version, and it changes the complexion of the story without excusing anything.
In June 2025, a federal judge ruled that destructively scanning books the company had lawfully purchased was fair use. The practice was tested in court and found lawful. This file alleges no crime by any company or any person, because on the question actually litigated, there was none.
That forces a better question than the one most posts are asking. If it was legal, why the codename? The answer is in the material itself, and it is not about law. The deleted ISBNdb page said it out loud: destroying two million books is not a headline that generates sympathy. The April 2024 memo reads like an organisation managing a reputational problem rather than concealing an unlawful one.
And a legitimate defence of the method exists, whether or not one finds it sufficient. Non-destructive scanning is what libraries do, and it is slow and costly because something has to turn every page. Cutting the spine makes throughput at the scale of millions possible at all. The destruction is a consequence of the method, not its objective. Nobody set out to reduce the world's book stock; they set out to move fast, and this is what moving fast costs.
Whether that trade is acceptable is a real argument, and the honest place to have it is here, on the merits of the method and what was lost, rather than on a false premise that a law was broken.
Why a six-month-old story detonated in July
The most interesting thing about this claim is its timing. The documents were public in January. The Washington Post reported them. Then nothing much. The explosion came in July, six months later.
What changed was not evidence but framing. January's coverage was about a copyright lawsuit and unsealed filings, which is a legal story. July's was about books being pulped, which is a story about a physical object with centuries of cultural weight behind it. The second needs no explanation of fair-use doctrine to land.
The quotable memo did the rest. “We don't want it to be known that we are working on this” requires no journalist to interpret it and no expert to validate it. Readers could evaluate it themselves, which is rare and which builds warranted trust.
And that warranted trust is what carried the unwarranted parts. Because the core was verifiable and verified, the additions rode along unexamined. This is the inverse of the usual failure: not a false claim that people believed, but a true claim whose true parts vouched for its false ones. A story that is mostly right is harder to correct than one that is wrong, because correcting it feels like defending the thing being criticised.
What is established, and what is not
Sorting this one is unusually straightforward, because the primary documents are public and anyone can read them.
Established: a confidential programme called Project Panama existed; its stated aim was to destructively scan all the books in the world; the codename existed because the company did not want the work known; books were bought in bulk from booksellers in several countries; millions had been purchased by October 2024; scanning required cutting the spines; and a federal judge found the practice to be fair use.
Not established: that other AI companies ran comparable programmes; that the books targeted were rare or collectible in the sense readers took from the word; that any book-data broker actually brokered the purchases; and whether the programme is still running, a question Snopes put to the company without receiving an answer.
So the rating is substantiated, applied to the claim as the documents support it rather than as the posts stated it. That is a rare verdict on this site, and it is worth being precise about what it means: it means the described conduct happened. It does not mean anyone acted unlawfully, and the court that examined the question found otherwise.
What is left is not a mystery but a question of judgement, and it is a better question than the one the posts asked. Millions of books were bought, cut apart, and not reassembled, lawfully, to build a machine, by an organisation that wrote down that it would rather this not be known. Nothing in the file resolves whether that was worth doing. It does establish, unusually clearly, exactly what was done.
What's still unexplained
- Whether Project Panama is still running is unknown. Snopes asked Anthropic directly whether the project was ongoing and reported no response by publication.
- What 'less common' meant operationally is the most consequential open question, because it separates a data-coverage strategy from the destruction of scarce cultural objects. The documents support the phrase but not an interpretation of it.
- Whether other AI companies ran comparable programmes remains unconfirmed. The existence of brokers marketing bulk physical-book acquisition suggests industry-wide demand, but no other company's internal records have been unsealed.
- The unresolved policy question is whether a fair-use finding is the right frame for a practice that consumes physical objects. The court addressed copying, not disposal, and no legal regime currently treats the destruction of a lawfully owned book as a distinct harm.
Point by point
The claim: An AI company was secretly destroying books to train a chatbot.
What the record shows: This is established, and by an unusually strong class of evidence: the company's own internal documents, unsealed on a judge's order and public since January 2026. The April 2024 memo states the purpose in one sentence, that Project Panama is the effort to destructively scan all the books in the world, and states the reason for the codename in another, that the company did not want the work known. Bulk purchasing is documented in correspondence with booksellers. Scanning at industrial speed requires cutting the spine off a book so the pages can be fed through a sheet-feeder, and the pages are not rebound afterwards. This is not a case where a claim outran its evidence. The evidence arrived first, in January, and the claim caught up with it six months later.
The claim: They were targeting rare and antique books specifically.
What the record shows: This is the part the record does not settle, and it is the part that gave the story its emotional charge. What the documents show is an August 2024 email in which the executive leading the project agreed that 'less common books are a great place to start', which supports a focus on scarcer titles but does not define it. A 211-page process document uses a link to a Book Industry Study Group category called Antiques and Collectibles as an example of categorisation, and it is not clear whether that indicates targeting antique books or simply books about antiques, which is a genuinely different thing. Snopes asked the company for clarity and had not received a response. 'Less common' in a data-acquisition context may mean titles poorly represented in existing digital corpora rather than physically scarce objects. Nobody has established that first editions or collectible copies were pulped, and this file does not assert it.
The claim: AI companies, plural, have all been doing this.
What the record shows: Only one is documented. Snopes rated the overall rumour Mostly True precisely because the plural could not be confirmed: it was not possible to establish whether other AI companies have carried out comparable projects. The 404 Media reporting that triggered the viral wave concerned a broker marketing bulk physical-book acquisition to AI developers generally, which is evidence that demand existed across the industry, not that specific competitors ran destructive-scanning programmes. The generalisation from one documented programme to an industry-wide practice is the single largest gap between the posts and the record.
The claim: A book-data company was brokering the destruction and then covered it up when caught.
What the record shows: ISBNdb's account and its deleted page do not sit comfortably together, and the honest summary is that the record is contested. A now-removed page on the company's blog described Anthropic acquiring, scanning and destroying books and included the line about two million books returned to pulp, along with a candid remark that 'AI company destroys two million books' is not a headline that generates sympathy. After the coverage, ISBNdb stated it has never purchased, scanned or sold a book for AI training or anything else, that it does not and never has trained AI models, and that the page was a test of market interest for a service that was never brought to life. Both things can be true: a company can market a service it never delivers. What the deleted page establishes is that someone in the book-data business thought there was a market. It does not establish that this company brokered any of it.
The claim: This was illegal, and that is why it was kept secret.
What the record shows: A federal judge ruled in June 2025 that destructively scanning books the company had lawfully bought was fair use. On the question the court actually decided, the practice was lawful, and this file makes no allegation of criminality against any company or person. That leaves the secrecy needing a different explanation than illegality, and the documents supply one themselves: the reputational problem. The deleted ISBNdb page put it plainly in noting that destroying two million books is not a sympathetic headline. Something can be entirely legal and still be the kind of thing an organisation would rather not read about itself, and the April 2024 memo reads much more like a company managing that than one hiding a crime.
The claim: Destroying the physical copy is the point. They could have scanned them without wrecking them.
What the record shows: Non-destructive scanning exists and is what libraries use, but it is slow and expensive per volume, because a human or a robotic arm must turn every page. Cutting the spine and feeding loose sheets through a document scanner is dramatically faster and cheaper, which is what makes it viable at a scale of millions. The choice is about throughput, not about wanting the books gone. That distinction matters for how the practice should be judged: the loss of the physical copies is a consequence of the method rather than its purpose, which is a fair thing to criticise but a different criticism from the one most posts are making.
Timeline
- 2024-04-13An internal Anthropic memo defines the programme: 'Project Panama is our effort to destructively scan all the books in the world.' It explains the codename directly, saying a soft codename is used 'because we don't want it to be known that we are working on this', that the document is visible to all employees but should not be discussed in public areas, and that the work should not be shared with anyone outside the company. The project is led by Tom Turvey, who had previously worked at Google on partnerships including Google Books.
- 2024-08-27An email exchange begins between Turvey and Wonder Book, a firm that sells books in bulk. Turvey confirms that 'less common books are a great place to start'. Separate correspondence shows the company approaching booksellers and publishers in other languages, including the Spanish publisher Oceano.
- 2024-10An internal memo indicates the company has already purchased millions of books. The exact figure is redacted in the released version, but the surviving 'M' suffix establishes the order of magnitude. A separate 211-page partially redacted document describes the process, including how books are categorised.
- 2024A group of authors sues Anthropic, alleging it trained its models on their work without permission and that this was copyright infringement. The company argues the use was fair use. The case becomes the vehicle through which the internal records eventually reach the public.
- 2025-06A federal judge rules on the fair-use question. Destructively scanning books the company had lawfully purchased is found to be fair use. This is the part most often missing from the viral version: the conduct was tested in court and held to be lawful.
- 2026-01-27The judge having ordered the documents unsealed, the internal records become public and are reported by major outlets, including The Washington Post under the headline that the company destructively scanned millions of books to build Claude. At this stage the story is covered but does not go viral.
- 2026-07-21The investigative outlet 404 Media publishes a piece on AI companies buying old books because they are free of AI-generated text, identifying a book-data company, ISBNdb, as having marketed high-volume physical-book acquisition to AI developers. A since-removed ISBNdb blog page had described the scale in memorable terms: 'Two million books. Bought, read by machines, and returned to pulp.' ISBNdb later states that it has never purchased, scanned or sold a book for AI training or anything else, that it does not train AI models, and that the page was a test of market interest for a service never launched.
- 2026-07-25 to 2026-07-29The story spreads. Futurism publishes on the scale of the practice, Novara Media follows, and posts on Instagram, Facebook and Reddit reach a much larger audience than the January reporting had, many quoting the 'we don't want it to be known' line directly from the unsealed memo.
- 2026-08-03Snopes publishes a fact check rating the claim Mostly True: confirming the destructive scanning, the bulk purchasing, the focus on less common books and the confidential codename, while recording as undetermined whether other AI companies did the same and what 'less common' meant for actual rarity. Snopes says it contacted Anthropic to ask whether the project is ongoing and will update if the company responds.
Proven. The central claim is documented, and the documentation is the company's own. Internal records unsealed by a judge in a copyright lawsuit and made public in January 2026 show that Anthropic ran a confidential programme it called Project Panama, described in an April 2024 memo as its effort to destructively scan all the books in the world. The same memo explains the codename: the company did not want it known that it was working on this, and told employees not to discuss it outside the company. Books were bought in bulk, sliced at the spine so pages could be fed through a scanner, and not reassembled. An internal memo from October 2024 indicates millions of books had already been purchased. In an August 2024 exchange with a bulk bookseller, the executive leading the project agreed that less common books were a good place to start. Two parts of the viral version are not established and this file does not assert them: whether other AI companies ran comparable programmes could not be confirmed, and what less common meant in practice, as against genuinely rare or collectible titles, is unclear. One point is frequently missed and belongs in any honest account: a federal judge ruled that destructively scanning lawfully purchased books was fair use, so this was found to be legal. The rating here is that the described conduct occurred, not that any law was broken.
Reviewed by The Conspiratory Editors · Last reviewed August 4, 2026 · How we rate
Common questions
Is Project Panama true?
Proven. The central claim is documented, and the documentation is the company's own. Internal records unsealed by a judge in a copyright lawsuit and made public in January 2026 show that Anthropic ran a confidential programme it called Project Panama, described in an April 2024 memo as its effort to destructively scan all the books in the world. The same memo explains the codename: the company did not want it known that it was working on this, and told employees not to discuss it outside the company. Books were bought in bulk, sliced at the spine so pages could be fed through a scanner, and not reassembled. An internal memo from October 2024 indicates millions of books had already been purchased. In an August 2024 exchange with a bulk bookseller, the executive leading the project agreed that less common books were a good place to start. Two parts of the viral version are not established and this file does not assert them: whether other AI companies ran comparable programmes could not be confirmed, and what less common meant in practice, as against genuinely rare or collectible titles, is unclear. One point is frequently missed and belongs in any honest account: a federal judge ruled that destructively scanning lawfully purchased books was fair use, so this was found to be legal. The rating here is that the described conduct occurred, not that any law was broken.
What is Project Panama?
In July 2026 a claim spread across Instagram, Facebook and Reddit that AI companies were buying obscure and rare physical books in bulk, cutting them apart to scan them, and destroying them to build chatbots, under an internal programme whose own documentation said the work should be kept quiet. Unusually for something that travels that…
What does the evidence show?
This is established, and by an unusually strong class of evidence: the company's own internal documents, unsealed on a judge's order and public since January 2026. The April 2024 memo states the purpose in one sentence, that Project Panama is the effort to destructively scan all the books in the world, and states the reason for the coden…
Why do people believe it?
The strongest line in the story is a direct quotation from the company's own memo. 'We don't want it to be known that we are working on this' needs no interpretation, no expert, and no trust in a journalist. People believed it because they could read it themselves.
What is still unresolved?
Whether Project Panama is still running is unknown. Snopes asked Anthropic directly whether the project was ongoing and reported no response by publication.
Sources
- 1.Are AI companies scanning and destroying millions of books, including rare titles?, Snopes (2026)
- 2.AI Companies Are Buying Tons of Old Books Because They're Free of AI Slop, 404 Media (2026)
- 3.AI Firms Are Buying up Old Books, Then Scanning and Destroying Them, Novara Media (2026)
- 4.Fact Check: At least one AI company is scanning and destroying millions of books, including 'less common' titles, Snopes, syndicated via Yahoo News (2026)
- 5.Bartz v. Anthropic, Wikipedia (2026)
- 6.Antiques and Collectibles (subject category), Book Industry Study Group (2026)
Embed this verdict on your site
Paste this snippet to show our sourced verdict as a small card, with a link back to the full case file. Free to use.
<iframe src="https://theconspiratory.com/embed/project-panama-book-scanning" title="The Conspiratory verdict" width="520" height="190" style="border:0;max-width:100%" loading="lazy"></iframe>Help us investigate
This is a living case file. If you spot an error or know evidence we missed, tell us, and weigh in on where you land.
Where do you land?
Cast your read on this one.
Comments
Add your take. Comments are read and approved by a human before they appear, so keep it on topic and civil. Please do not accuse named, living people of crimes.