AI's 'Book Burning': Why Companies Are Destroying Millions of Physical Books to Train Their Models

Some of the most valuable AI models on earth were built, in part, on a quiet act of destruction: companies bought millions of physical books, sliced the bindings off, scanned the pages, and threw the originals away. This week that fact went viral all over again — reframed by critics as "AI book burning" — as people confronted the strange reality that teaching a machine to read has meant physically destroying the books it learned from.

The physical destruction is documented. The legal question is narrower: the court considered training, converting purchased copies and retaining a library of pirated copies separately.

What actually happened

The clearest documented case is Anthropic, the company behind Claude. To assemble a massive, high-quality library of text to train its models, Anthropic didn't just scrape the web. It bought millions of physical print books and ran them through a process the industry politely calls "destructive scanning": the spines are cut off, the loose pages are fed through a scanner, the text is digitized, and the physical book — now a pile of shredded paper — is discarded.

This wasn't a rogue side project. Court filings revealed Anthropic hired an executive who had previously helped run Google's book-scanning program specifically to acquire books at enormous scale. The goal was blunt: get as many books as possible into the training set.

The twist: they destroyed the books on purpose, for legal reasons

Here is the counterintuitive part. Anthropic didn't shred books out of carelessness — it did so because destroying them was the safer legal path.

Buying a print copy and retaining a downloaded pirated copy raised different issues in this litigation. The court assessed each use separately rather than declaring a general rule for all AI training.

In the court’s account of its fair-use rulings, training on the works at issue and one-for-one conversion of purchased print copies for a searchable internal library were fair uses. Downloading and retaining pirated copies in the central library was not justified by fair use on the summary-judgment record. Those are distinct findings; a settlement is not a trial verdict on every allegation.

The ruling does not establish that buying and destroying a book removes every copyright risk. Its findings depend on the particular uses and record before that court.

Why people are calling it "book burning"

The phrase is emotionally charged — and technically inaccurate; nothing is being set on fire. But it captures a real unease. Millions of books were turned into shredded paper so their contents could live inside a chatbot. For most bulk-bought used paperbacks, that's arguably no great loss — the same text exists in countless other copies.

The sharper worry, and the reason this resurfaced now, is what happens at the edges: whether rare, antique, or out-of-print editions — books where each surviving copy actually matters — could get swept into industrial-scale buying and shredding. Once a scarce edition is cut up and tossed, that physical object is gone for good, even if its words survive as data. Critics argue that treating all books as interchangeable "training fuel" misunderstands what a book is. Defenders counter that the companies bought ordinary copies in bulk, that the text is preserved, and that a court blessed the practice.

Even rival AI leaders have used the moment to score points. Elon Musk publicly signaled that his own AI efforts wouldn't be shredding rare books to train models — a pointed jab at Anthropic that kept the story burning across social media.

The bigger picture

Strip away the "book burning" framing and you're left with something that says a lot about the AI era: the demand for training data is now so intense, and the copyright rules so unsettled, that destroying physical books became the legally cautious option. It's a vivid example of how AI's hunger for human-made content keeps colliding with the systems — copyright law, libraries, publishing — built to preserve that content in the first place. The books-into-data pipeline is unlikely to stop; the fight now is over where its limits should be.

Frequently asked questions

Did AI companies really destroy millions of books?

Yes. Court records in the Bartz v. Anthropic case established that Anthropic bought millions of physical books and destructively scanned them — cutting the bindings, scanning the pages, and discarding the physical copies — to build a training dataset for Claude. The practice is real and documented.

Is it actually "book burning"?

Not literally — no books are set on fire. "Book burning" is an emotionally charged label critics use for the mass destruction of physical books. The books are cut apart and shredded, and their text is preserved digitally in the training data.

Why would a company destroy books it paid for?

The court accepted the one-for-one conversion of purchased books in this case. Destruction of the original was part of that factual record, not a universal legal requirement for every scanning project.

Were rare or irreplaceable books destroyed?

The documented practice centered on buying ordinary used books in bulk, where the same text exists in many copies. The current concern — and the reason the story went viral again — is whether rare, antique, or out-of-print editions could be caught up in industrial-scale buying and shredding, since those physical objects can't be replaced once destroyed. That specific worry is the contested part of the debate, not a settled fact.

Is this legal?

This case does not give a blanket answer for every dataset or training project. Its separate findings about training, purchased-copy conversion and pirated-library retention should not be collapsed into “all training is legal” or “all training on pirated works was ruled illegal.”

Reporting based on the public record of the Bartz v. Anthropic copyright case and Judge William Alsup's 2025 fair-use ruling, plus coverage by Tom's Hardware, Futurism, Decrypt and others. The Bot Post will update this story as the copyright landscape develops.

Editorial correction, October 8, 2026: Clarified the distinct fair-use findings and removed language implying that destructive scanning eliminates legal risk.