A Book Broker Told AI Clients "The Optics Problem Is Real." Nine Days Later the Page Was Gone.
Bot Mutiny |
ISBNdb marketed bulk print-book acquisition to AI labs and warned them how the destruction would look in a headline. After 404 Media published the pitch, the company deleted the pages and said it had never bought, scanned or sold a book.
The sentence that did the damage was on ISBNdb's own marketing page, and it was written for AI companies rather than about them.
"The optics problem is real," the page said, before spelling out the headline the company expected its clients to want to avoid: an AI firm destroying two million books.
That page is gone now.
What ISBNdb was selling
ISBNdb describes itself as the world's largest book database. Alongside that, it was offering high-volume acquisition: sourcing printed books in bulk, from libraries, retailers and estate sales, in orders it said could run to a million titles, for AI companies to turn into training data.
The pitch, reported by Emanuel Maiberg at 404 Media on July 21, rested on a genuinely interesting argument about contamination. So much of the open web is now machine-written that models trained on it risk feeding on their own output. A book printed before 2022 cannot contain text generated by a large language model, because the models did not exist. The company's material described books as dense, edited and authoritative in a way no web crawl reproduces, and argued that a pre-2022 print run is a fixed human record nobody can quietly rewrite.
There was a second argument about poisoning. Writers have begun lacing text with characters that read normally to a person and corrupt a model, using tools in the family of Nightshade. ISBNdb's own blog pointed at Anthropic research suggesting a few hundred crafted documents can plant a backdoor in a corpus of trillions of tokens. Books printed before any of that existed sidestep it.
Both arguments are sound. Neither is the story.
The walkback
On July 30, nine days after the 404 Media piece, ISBNdb added a note to its homepage and an update to its news page. The acquisition pages had been deleted.
The company's position is now that it has never purchased, scanned, or sold a book, for AI training or for anything else, and that the material was a test of market interest.
Both things can be true at once. A company can publish a marketing page describing a service it has not yet performed. That is what a test of market interest is. What the page shows is not that books were destroyed. It is what the company believed its intended customers would find persuasive, and it thought they would be persuaded by reassurance about how the destruction would look.
One of the removed pages argued that buying paper books at scale from the secondary market takes nothing from creators, because those copies have already discharged their financial obligation to the people who wrote them. That is a defensible reading of first-sale doctrine. It is also an argument written for someone already thinking about the objection.
Why the pitch worked
The commercial logic did not come from nowhere. It came from a court.
In June 2025, in Bartz v. Anthropic, Judge William Alsup of the Northern District of California held that digitizing legally purchased print books and training a large language model on the results was transformative fair use. Part of his reasoning was that the process replaced the print copy with a digital one rather than adding copies or redistributing them. Anthropic later settled separate claims involving pirated digital books for $1.5 billion.
Read that holding as a buyer and the incentive is visible. If the legal safety of the scan rests partly on the original being replaced rather than duplicated, then not keeping the paper copy is not merely convenient. It is closer to the shape of the thing the court approved.
Filings unsealed in the same litigation described Anthropic's internal program, which an internal planning document called an effort to "destructively scan all the books in the world." The same document noted the project used a soft codename. In February 2024 the company had hired Tom Turvey, formerly head of partnerships for Google's book-scanning project. Court records describe service providers stripping bindings, trimming pages, scanning them and discarding the paper.
What the record establishes, and what it does not
This is where most of the coverage has collapsed two things into one, and the distinction is the whole point.
The record establishes that ISBNdb marketed bulk print-book acquisition to AI developers, that its materials told prospective clients the optics problem was real, that a bulk bookseller told 404 Media that large orders have surged in the past year and that uncommon books are being pulped, that the pages came down nine days after publication, and that the company now denies ever having bought, scanned or sold a book.
Separately, the record establishes that Anthropic ran a destructive scanning program, in its own words, and that a federal judge found the underlying practice to be fair use.
These are two different source chains, and nothing published so far connects them. Neither ISBNdb's marketing nor 404 Media's reporting identifies an AI buyer behind a completed ISBNdb order, or shows that any such order ended in a book being destroyed. Writing this as "ISBNdb shredded books for Anthropic" would be false, and it is roughly what a reader would take away from several of the aggregated versions now circulating.
The record also does not establish scarcity. ISBNdb's material used the words rare and out of print. Neither says anything about how many copies of a title exist. A book can be hard to buy without being unique, and a copy can be replaceable even when its edition is uncommon. Whether anything irreplaceable has been pulped is unknown, and the people best placed to know are the booksellers filling the orders.
What it does establish is smaller and harder to argue with. A court ruling made not keeping the paper the safer path. A broker built a product on that reading and wrote its customers a paragraph about the headline risk. And the paragraph came down nine days after somebody printed it.
Sources: 404 Media, "AI Companies Are Buying Tons of Old Books Because They're Free of AI Slop," July 21, 2026, and "Company Offering Printed Books to Train AI Stops After 404 Media Coverage," July 31, 2026, both by Emanuel Maiberg; ISBNdb's now-removed acquisition pages and its July 30 homepage note; Bartz v. Anthropic, N.D. Cal., order of June 23, 2025 (Alsup, J.); filings unsealed in the same matter describing Project Panama; Washington Post reporting on Anthropic's book purchases.