Anthropic Book Copyright Case: Myths vs the Real Ruling

What did the Anthropic book copyright ruling actually say?

The Anthropic book copyright ruling split the case cleanly in two: training was lawful, piracy was not. In June 2025, US District Judge William Alsup ruled on summary judgment in Bartz v. Anthropic that using legally acquired books to train the Claude large language model was “quintessentially transformative” and protected as fair use under Section 107 of the US Copyright Act.

The second half of the ruling is the half the headlines tend to drop. Alsup held that Anthropic’s downloading and permanent storage of pirated books from shadow libraries such as Library Genesis and Pirate Library Mirror was not fair use, and looked like straightforward infringement. As reported by NPR and Reuters, the judge wrote that Anthropic had no entitlement to build a central library out of pirated copies, and that later buying some of those titles did not wipe out the original wrong.

That piracy exposure is what drove the settlement. Rather than take the piracy claim to a damages trial, Anthropic agreed in 2025 to pay $1.5 billion, which works out at roughly $3,000 per book across an estimated 500,000 covered titles.

Final approval came in July 2026. So the accurate one-line summary is this: The training was cleared, the piracy was paid for.

The distinction between how the data was obtained and what was done with it runs through everything below. Keep it in mind for the next myth, which leans heavily on ignoring it.

Is the Anthropic case really like Orwell’s 1984 or Bradbury’s Fahrenheit 451?

No. The comparison to 1984 and Fahrenheit 451 is emotionally satisfying and factually empty.

Both novels are about state power destroying ideas: Orwell’s Ministry of Truth rewrites history and erases inconvenient facts, while Bradbury’s firemen burn books to stop people thinking. Neither maps onto a copyright dispute about how a private company sourced training data.

The core of both books is suppression of knowledge. In the Anthropic book copyright case, no idea was censored, no title was banned, and no reader lost access to anything. Every book involved remains in print, in libraries, and on shelves. A model learning statistical patterns from text is close to the opposite of a regime deleting text so nobody can read it. If anything, the technology makes the contents of books more findable, not less.

There is a fair criticism buried under the bad metaphor, and it deserves stating plainly rather than dressing up in dystopian costume.

Authors argued that Anthropic sought to profit from their work without permission or payment, and on the piracy point the court agreed and the company paid. That is a real grievance about consent and compensation. It is not thought control, and reaching for Orwell or Bradbury actually weakens the genuine argument by making it sound hysterical.

The book-burning image gets its real power from the idea that physical books were destroyed. That part is not invented, so it is worth looking at honestly, alongside what the publishing industry already does with books every year.

Did Anthropic “destroy books”, and how many does the UK pulp anyway?

Anthropic did physically destroy books, and the scale is real, but the framing still matters. To build a searchable digital library, Anthropic ran an internal operation it codenamed Project Panama, revealed through court filings unsealed in January 2026 and reported by the Washington Post.

Vendor proposals referenced converting between 500,000 and two million books over six months. The process was industrial: a hydraulic cutter removed the spines, high-speed scanners captured every page, and the paper was then pulped & recyled. Judge Alsup found this scan-and-discard method, applied to lawfully purchased books, fell within fair use, on the reasoning that each physical copy was simply replaced by a single digital one under the first-sale doctrine.

There is a genuine concern here that deserves stating without exaggeration. Follow-up reporting by 404 Media in July 2026 found that a market has grown up around sourcing physical books for AI training in bulk, often through brokers using confidentiality agreements, drawing on second-hand, overseas and foreign-language stock rather than fresh in-print copies.

One bookseller told 404 Media they had mixed feelings, welcoming the clear-out of unsellable inventory but disliking that:

“uncommon books are being pulped”.

The worry that an out-of-print title with few surviving copies could go through a cutter is legitimate. What has not been shown, as a media fact-check by MLQ News noted in July 2026, is a single verified rare title that an AI company bought and destroyed. The concern is real; the proof of catastrophic loss is not yet on the record.

Here is the context almost nobody supplies, and it is worth putting in real numbers rather than percentages. The UK buys roughly 200 million print books a year, going by Nielsen BookScan’s tracked sales.

On top of that figure, the trade ships far more copies to shops than it ever sells, and the unsold ones come back. Industry analysis puts average returns at around 20 to 25% of the copies distributed, with mass-market paperbacks reported as high as 48%.

Put plainly, for every hundred books that sell, another twenty-five to thirty-five are printed, shipped, and sent straight back.

That works out at somewhere between 40 and 65 million returned books in the UK every year, and a large share of those are pulped rather than resold, because warehousing them costs more than destroying them. To put that side by side with the case at hand: even the low end of that annual pulping, 40 million books, is between twenty and eighty times the entire 500,000 to two million that Anthropic scanned across its whole Project Panama operation. The routine, unremarked destruction dwarfs the thing people are furious about.

This is not a fringe practice. Amazon alone was shown by an ITV News investigation to be marking well over 100,000 items a week for destruction in a single UK warehouse, books among them. Much of the outrage, then, is in the selective framing. Quietly pulping a quarter of a print run for warehouse economics passes without comment, while scanning bought books to make their contents searchable gets compared to totalitarian censorship.

The out-of-print worry is worth taking seriously, but it applies just as much to the routine pulping the industry already does, and that half of the ledger rarely gets a headline.

None of this means the copyright concern is fake. It means the specific “they destroyed books” line proves far less than it appears to, which brings us to the biggest legal misreading of all.

Did the judge rule that AI training is copyright infringement?

No, the judge ruled close to the opposite on training, and that misreading is the most consequential myth in circulation. On the specific question of using copyrighted books to train an AI model, Judge Alsup sided with Anthropic, finding the training use transformative and protected as fair use. The infringement finding was about piracy, meaning the downloading and hoarding of stolen copies, not about the act of training itself.

Fair use is a US legal doctrine, decided case by case, that can permit copying without permission depending on factors such as the purpose of the use and its effect on the market for the original. Because it is fact-specific, this ruling does not make all AI training automatically legal everywhere.

Two further limits matter for anyone quoting it. First, because Anthropic settled rather than letting the case reach an appeals court, the fair-use ruling is a single district-court decision and never became binding precedent. Second, this is US law, and international copyright treaties do not turn it into a global verdict.

That treaty point is worth spelling out, because it is where the argument usually gets muddled. Under the Berne Convention, which both the US and the UK have signed, a UK author’s work is automatically protected in the United States to the same standard a US author’s work receives. That principle, known as national treatment, is why British writers were covered by the class action and could claim against Anthropic in a US court. What Berne does not do is export any one country’s exceptions. It runs on territoriality: the governing law is the law of the country where the use takes place, not the author’s home law. Anthropic’s scanning and training happened in the United States, so US fair use governed whether those acts were lawful.

The practical consequence for a British business is the opposite of reassuring.

The UK has no broad fair-use defence of its own. It has narrow, purpose-specific exceptions, plus a text-and-data-mining exception currently limited to non-commercial research, and the scope of that exception is exactly what the live UK government consultation on AI and copyright is contesting as of 2026. There is no equivalent doctrine for a US fair-use finding to carry across, so the same conduct judged in a UK court would start from a very different place.

The table below lays out what the case did and did not decide, because the gap between the two is exactly where the myths breed.