#isbn

There is a real story here. At least one AI company has purchased and destroyed millions of physical books after scanning them. Booksellers have also reported unusual bulk orders that they suspect may be connected to AI training.

But “AI companies are destroying books” and “AI companies are destroying antique books at incredible scale” are not the same claim.

Let us separate what has been documented from what has been assumed.

What actually happened

In 2025, court documents revealed that Anthropic had spent millions of dollars purchasing millions of physical books, often second-hand.

The company’s contractors removed the bindings, cut the pages to a uniform size, scanned them, and discarded the paper copies. Each physical book was replaced with a searchable digital copy in Anthropic’s internal research library.

That part is true.

Anthropic really did buy and destroy millions of books. This is not speculation based on unusual orders or anonymous sources. It is described directly in a federal court ruling.

However, the ruling dealt with several legally distinct actions.

The court found that using books to train Anthropic’s language models was fair use under the circumstances presented. It also found that replacing each legally purchased print copy with one internal digital copy was fair use because the original was destroyed, the total number of copies did not increase, and the digital replacement was not distributed outside the company.

The court did not find that Anthropic was entitled to download and permanently retain millions of pirated books merely because some might eventually be used for AI training. Those piracy claims were later covered by a US$1.5-billion settlement that received final approval in July 2026.

So, yes: lawfully purchased books were destructively scanned.

But no: the ruling did not give AI companies a universal licence to acquire books however they please.

Where did the “antique books” claim come from?

The recent panic began with reports about unusually large and apparently random book orders.

One bookseller told 404 Media that his sales had suddenly risen from fewer than twenty books a week to hundreds. The orders showed little apparent interest in subject, author, genre, or even reasonable pricing. The books reportedly had ISBNs, leading the seller to suspect that they were being purchased in bulk for some form of automated processing.

The seller also carried foreign-language, uncommon, and out-of-print books. He therefore worried that some of the copies might be scarce and could be destroyed after scanning.

That concern is reasonable.

It is not proof that antique or irreplaceable books are being destroyed at an “incredible scale.”

Even Futurism’s article acknowledges that booksellers generally do not know who placed the orders or what happened to the books afterwards. The buyers may be AI companies, but the sellers are largely inferring that from the unusual purchasing patterns.

Some uncommon books may be disappearing into private scanning operations.

That possibility deserves investigation.

It should not be reported as a proven worldwide assault on antiquarian collections.

An ISBN is not a rarity certificate

The fact that the reported bulk orders focused on ISBN-listed books is also important.

An ISBN is a commercial product identifier. It identifies a particular title, edition, publisher, and format so that booksellers, libraries, distributors, and other participants in the book trade can catalogue and order it.

It does not tell us whether an individual copy is rare, valuable, historically important, or one of thousands gathering dust in second-hand shops.

ISBNs were issued as ten-digit numbers until the end of 2006 and as thirteen-digit numbers from 2007 onward. Older ten-digit ISBNs can also be converted into thirteen-digit equivalents, so a modern database displaying an ISBN-13 does not necessarily mean that the book itself was published after 2007.

A book can have an ISBN and still be scarce or out of print.

It can also be an extremely common paperback that nobody has wanted since 1993.

“ISBN-listed,” “out of print,” “rare,” and “antique” are not interchangeable descriptions.

The alleged middleman has now walked back its claims

Much of the recent reporting focused on ISBNdb, a book-metadata company whose website advertised bulk book-acquisition services for AI developers.

Its marketing pages claimed it could source between 1,000 and one million books per order. They promoted books published before the recent proliferation of generative AI as a supply of supposedly uncontaminated human-written material. They also discussed confidentiality and acknowledged that headlines about AI companies destroying millions of books would create an “optics problem.”

Those pages certainly existed.

But after 404 Media reported on them, ISBNdb removed the material and issued a denial. The company now says the pages were only a test of potential market interest, that the service was never launched, and that it has never purchased, scanned, or sold a physical book for AI training.

That denial does not prove that no other company is sourcing books for AI laboratories. It does mean that we should not cite ISBNdb’s abandoned marketing pitch as proof that such purchases are already occurring through that company at the advertised scale.

A company advertised a possible service.

Booksellers noticed strange orders.

Some sellers suspect AI buyers.

Those facts justify questions.

They do not justify presenting every suspicion as a confirmed industrial operation.

Book destruction did not begin with AI

There is another piece of context missing from much of the outrage: the book trade already destroys books routinely.

Bookstores commonly return unsold stock to publishers or wholesalers. Some returned copies are resold or remaindered. Others are pulped, recycled, or otherwise destroyed because storing, shipping, inspecting, and redistributing them would cost more than they are likely to earn.

An industry report on international bookselling found that many returned books are pulped, while also noting how little transparency exists concerning the number destroyed each year. The International Publishers Association describes pulping unwanted titles as an ordinary outcome of the returns system.

The same thing happens in the second-hand economy.

Charity and thrift shops cannot keep every donated book indefinitely. Shelf and warehouse space are finite. The Salvation Army has openly explained that books unsuitable for resale may be sent to cardboard and paper-pulp recyclers.

None of this makes Anthropic’s destruction of books admirable.

It does expose a glaring double standard.

When publishers pulp unsold stock because they printed more copies than consumers wanted, it is treated as an obscure and unfortunate feature of the publishing business.

When an AI company buys used books, scans them, and destroys the originals, it becomes evidence that technology is devouring human culture.

The physical outcome may be the same.

The symbolism is different.

Why the symbolism matters

Books are not perceived as ordinary consumer goods, even when the industry treats them that way.

A warehouse full of unsold paperbacks quietly sent for recycling rarely produces viral headlines. An industrial machine cutting the spines from books so that an AI company can extract their contents is much more emotionally potent.

It looks like a machine literally consuming literature.

That reaction is understandable. Books can be cultural artefacts, personal objects, historical records, and vessels of human expression.

But the emotional power of an image does not relieve journalists of the obligation to report accurately.

If the concern is the destruction of culturally important books, then the relevant questions are whether the books are genuinely scarce, whether other copies have been preserved, and whether the resulting scans will remain accessible.

Simply calling them “antique” does not answer any of those questions.

Does buying a book make AI training legal?

Lawfully purchasing the books matters, but it is not a magic permission slip.

In the United States, one federal court ruled that Anthropic’s training use was fair use under the facts before it. The court separately ruled that replacing legally purchased print copies with one-for-one internal digital versions was fair use.

Those were related but distinct legal conclusions.

Other countries do not necessarily use the American doctrine of fair use. Some instead have specific exceptions for text and data mining.

European Union law permits text and data mining of lawfully accessible material under certain conditions, although rightsholders may reserve some uses. The United Kingdom currently permits text and data mining without additional permission only for non-commercial research where the researcher already has lawful access.

The accurate conclusion is therefore not:

Buying a book automatically makes commercial AI training legal everywhere.

It is:

Lawful acquisition can be an important part of whether computational analysis is permitted, but the answer depends on the jurisdiction, the purpose, and the exact use being made of the copies.

That may be less exciting than a sweeping declaration.

Copyright law has never been famous for its thrilling simplicity.

What deserves genuine scrutiny

There are still legitimate issues here.

Companies should be transparent about where their training material comes from and how it is acquired.

Books that are genuinely scarce, historically significant, uniquely annotated, or otherwise important should not casually disappear into inaccessible corporate archives.

Destructive scanning may also be wasteful when non-destructive methods are practical, particularly if the physical book would otherwise remain available to readers.

And a privately held scan is not the same thing as preservation. A book removed from circulation and converted into proprietary training data has not necessarily been preserved for the public.

Those concerns are serious enough without pretending that AI companies have already been proved to be systematically shredding the world’s antique treasures.

The truth does not need help from clickbait

The confirmed story is already strange and uncomfortable.

Anthropic legally purchased millions of physical books, destroyed them while making digital replacements, and used books from its larger library in developing language models.

Other booksellers have seen mysterious bulk orders that may be connected to similar projects.

Some uncommon books could be involved.

But there is currently no reliable evidence demonstrating that AI companies are destroying antique or irreplaceable books at an enormous global scale.

That conclusion was created by stretching a concern into a certainty, a suspicion into an industry-wide fact, and “out of print” into “antique.”

Criticism of AI companies should be based on what they actually do.

When legitimate concerns are inflated into clickbait, the result is not stronger accountability. It is weaker journalism — and another reason for readers to distrust the next alarming headline, even when that one might be true.

Sources

Bartz v. Anthropic PBC: federal court order on AI training, purchased print books, destructive scanning, and pirated digital libraries. [Link 1] [Link 2]

Reuters: final approval of the US$1.5-billion Anthropic copyright settlement. [Link]

404 Media: original ISBNdb report and the company’s subsequent denial and removal of its book-sourcing pages. [Link 1] [Link 2]

Futurism: the article claiming that antique books are being destroyed at an incredible scale. [Link]

International ISBN Agency: what an ISBN identifies and the transition from ten to thirteen digits. [Link 1] [Link 2]

International Publishers Association and RISE Bookselling: returns, recycling, pulping, and the lack of reliable industry-wide totals. [Link 1] [Link 2]