The Books Were Real. The Rarity Is Unproven.

In July, a post started circulating that said AI companies are buying rare books in bulk, cutting them apart, and shredding what is left to train their models. It spread because it confirmed what a lot of people already suspected about how these systems were built. At a community discussion on AI this week, the story came up, and the honest answer was that it sounded plausible and nobody in the room could say whether it was true. Checking took a handful of searches and two court exhibits. The result is less tidy than the post and more uncomfortable than the rebuttals.

What the record confirms

Court filings unsealed in January in Bartz v. Anthropic include an internal planning memo, last updated April 13, 2024, that opens by calling Project Panama the company's effort to "destructively scan all the books in the world." The same memo explains the codename: Anthropic did not want it known that the work was happening. The Washington Post's reporting on the filings describes the mechanics. Anthropic bought books in lots of tens of thousands, a hydraulic machine cut off the spines, high-speed scanners processed the loose pages, and a recycler collected the remains. The engineering documents name Better World Books and World of Books as sources of used stock, and distributors such as Baker & Taylor as sources of new stock. An October 2024 memo shows millions of books already purchased. Snopes reviewed the documents and rated the broad claim Mostly True.

Where the rarity claim comes apart

The viral version said rare, obscure, antique. That part has a weaker trail. It starts with a blog page from ISBNdb, a book-data company, pitching AI labs on bulk book sourcing. ISBNdb now says the service never launched and that it never bought, scanned, or sold a book for anyone. 404 Media quoted the page, a Futurism headline said antique books, and by the time the claim reached Instagram it read "obscure and rare."

The evidence for any rarity at all is thin and ambiguous. In an August 2024 email to a bulk bookseller, Anthropic's program lead wrote that less common books were a good place to start, according to Snopes, which could not tell what that meant for how scarce any given book was. Anthropic says none of its programs buy and destroy rare or antiquarian books, and, asked to clarify, it separated collectibles from harder-to-find professional and reference titles. A 211-page engineering handoff document from the same case fits a narrower reading. It describes software built to avoid buying what the library already holds, matching titles and authors against the existing database and against each order, and in one order removing a further 16 percent as duplicates. The document links to a publishing category called "Antiques and Collectibles," but only as an example inside a database schema. Reading "less common" as "not already in our collection" is an inference, and the document does not state it. If it holds, the program prized uniqueness inside Anthropic's own library, which is a different thing from scarcity in the world.

One widely shared detail belongs to someone else entirely. Fortune reported that a Dutch antiquarian bookseller received a request for 3,001 titles from a company called 2077AI, with shipping to China. The titles were mostly academic books published in 2020 and 2021, and the email never mentioned AI. It points to bulk book buying by AI-linked firms, though 2077AI did not answer Fortune's questions about the purpose. It says nothing about rare books, and nothing about Anthropic.

So the sentence the evidence supports is this one: Anthropic ran a secret program that destroyed millions of purchased books, prioritized "less common" titles, and no public record shows which copies went to the recycler. The sentence the post wanted travels better. It is also the one a skeptic can dismantle in an afternoon.

The law and the instinct point in opposite directions

Judge William Alsup ruled in June 2025 that training on lawfully acquired books was fair use, and that buying a print book, scanning it, and discarding the original was too. Copyright law has no objection to destroying an object you own, and destroying the original left the number of copies in circulation unchanged. The ruling went the other way for more than seven million books Anthropic downloaded from pirate libraries. Anthropic settled that part for $1.5 billion, which received final approval on July 20, 2026 and works out to roughly $3,000 per book across more than 400,000 works.

Notice what that rewards. The conduct that outrages people on preservation grounds, cutting apart millions of books, is the conduct the court cleared. The conduct that cost $1.5 billion, downloading pirated files, never destroyed anything. The settlement covers those files and not the pulped books, so the law measured one harm and cleared the other. Authors sit in a third position. A new copy pays its author a royalty and a used copy does not, and the record shows vendors of both kinds. Nothing in the filings reviewed for this piece says how many purchased copies were new.

Using it anyway

The pushback on social media was about how these systems were trained, and it raises the obvious follow-up: if the tool was built this way, why keep using it? My answer, after some time with it, is that the discomfort is accurate and the use continues. Squaring those two takes more than the usual defenses. The books are already pulped, so one more user changes nothing, which proves too much and ignores that subscriptions fund the next round of acquisition. The benefit outweighs the harm, except that the benefit lands on the user and the debt is owed to authors. The way an adversarial user prompts the tool is cleaner than the way it was built, except that the model argues back using the same books however it is prompted. Each of those describes how the tool is used and says nothing about where it came from.

Three separate questions get blended here. Is the tool reliable? Was it built legitimately? Is it changing how its user thinks? Provenance does not answer the first, and reliability does not answer the second. The verification above worked because primary documents existed and anyone can open them. Where no paper trail exists, the same search returns articles citing each other, and the tool summarizes the circularity with confidence. The legitimacy question stays open: the knowledge a user searches with is authors' labor, so whatever truth comes back has a creditor. On the third question, a tool tuned toward agreement can launder a bias as easily as expose one, which is why the habit of demanding friction has to be maintained by the user instead of trusted to the product.

One more fact belongs in plain view. The research behind this piece ran through Claude, built by Anthropic, the company at the center of the story. The tool reported the damaging parts of the record, and it also made one claim about used-book sourcing that the court documents then contradicted. A user cannot tell from the inside whether a tool's candor is honesty or trained posture, so every claim above traces to a court record, a fact-check, or a news report that a reader can open without it.

Responsibility here is not evenly shared, even though the conclusion is. The company chose how the tool was built. The user chooses only what to do next. For this user that means continuing, buying the books, new and used, listening to audiobooks, teaching from books in the classroom, and saying all of it in public. Those habits are real and small. They do not repay the authors whose pirated books trained the model. That debt was settled in dollars between a company and a class, and nothing a professor does touches it. What remains missing is a stop condition, a specific, observable event that would end the use. This piece does not invent one. It leaves the gap open, because a position that cannot name what would change it is closer to a mood than a position.

Act as a skeptical fact-checker working claim by claim. Here is something
I saw or shared that I'm not sure is true:

[paste the claim, post, or headline]

Do four things. First, break it into separate factual claims. Second,
sort each claim into one of three bins: documented in a primary source
(a court filing, official record, or original document), reported but
not independently confirmed, or unsupported. For every documented claim,
name the exact source I should open to check it, and do not describe a
source you have not actually read. Third, trace where the most
exaggerated version of the claim first entered circulation, and what
changed at each retelling. Fourth, tell me which single claim, if it
turned out to be false, would collapse the whole story.

If your maker, or a company you depend on, is named in the claim, say so
before anything else and tell me which of your conclusions I must verify
without you. Don't soften any of this to make the claim sound more or
less alarming than the record supports.

Act as a hostile but fair interlocutor. Here is a tool I use and a
harm in how it was built:

[name the tool or practice, and what you know about the harm]

My current position, in two sentences, on why I keep using it:

[state your position honestly]

Do not validate it. First, tell me which of these four rationalizations
my position relies on: the harm is already done, the benefit to me
offsets it, the way I use it is cleaner than the way it was built, or
small repayments cancel the debt. Quote my own words back to me to show
where. Second, separate three questions I may be blending: whether the
tool is reliable, whether it was built legitimately, and whether it is
changing how I think. Third, ask me what specific evidence would make me
stop using it, and keep asking until I give a condition I could actually
observe. Ask one question at a time and wait for my answer before the
next one.