AI companies are buying rare books, scanning them and sending them to the shredder

A rare book bought from a second-hand seller carried an AirTag placed there by 404 Media investigators. The tracker ended at VGT3, a secret warehouse described as a site where Amazon scans and destroys books for artificial intelligence training.

At an Amazon warehouse in Las Vegas, a separate 404 Media investigation found employees handling bulk shipments, removing covers, scanning pages for AI training datasets and throwing away the paper. In these operations, workers use hydraulic-powered cutting machines to sever spines and box cutters to break apart bindings. The separated pages are scanned at speed before being discarded or recycled.

Millions of physical books are being digitised and destroyed for AI models. They include rare second-hand volumes. Some books Amazon destroyed had only a few physical copies left in circulation worldwide, and many rare books could vanish permanently through the practice.

AI companies have shown particular interest in rare and out-of-print titles, which are generally unavailable in digital form. For years, they have bought used and rare books in bulk from second-hand sellers, including works on philosophy, history, medicine and law.

Second-hand bookshops across the United States and Japan received mysterious orders through multiple anonymous accounts, with deliveries sent to logistics centres or secret warehouses. Sellers had long suspected AI companies were behind the purchases, though the destination and purpose remained unclear until VGT3 was discovered in August.

Anthropic’s internal Project Panama puts the scale in unusually blunt terms. The company began the operation in early 2024 to scan and destroy physical books at scale, acquiring hundreds of thousands to millions of them from wholesalers. Its planning documents set a target of converting anywhere from 500,000 to two million books over a six-month period.

One company memo said: “Project Panama is our effort to destructively scan all the books in the world. We don't want it to be known that we are working on this.”

Anthropic has said it does not destroy rare or antique books. Amazon has made no promise that it will avoid destroying rare or antique volumes. A bookseller estimated that the number of books destroyed could be in the tens of millions.

Alexis Gunderson, president of the National Writers Union, said scan-and-destroy operations trample on the rights of authors, illustrators, photographers, translators and publishers, and should alarm readers, booksellers, librarians, educators and researchers.

The physical destruction is one part of a widening dispute over how AI companies obtain the text used to train large language models. On 10 July 2026, Hachette Book Group, Cengage Learning, Elsevier and the author Scott Turow filed a class action against Google LLC, alleging that it wilfully copied millions of copyrighted books without permission to train its Gemini AI models.

The suit alleges that Google violated limited-use agreements. Material from Google Books, Google Play and Google Scholar had been supplied under agreements that did not authorise AI training. Google’s training copies also came from material it already held, unauthorised web scrapes that included paywalled content, and pirate “shadow library” sources.

An internal Google document warned that training on copyrighted books could be “highly problematic for Google” and expose the company to “$10Bs–$100Bs in potential fines”. As of October 2026, Google had neither responded in court nor commented on the allegation that it stripped copyright-management information from training material in breach of the Digital Millennium Copyright Act.

There are at least 15 author-driven lawsuits against AI companies, alongside roughly 10 coordinated suits brought by publishers. Turow’s case sits within that broader legal fight, while parts of the publishing industry are taking a different route.

HarperCollins became the first of the Big Five publishers to disclose publicly that it had licensed content for AI training. Its three-year agreement pays $5,000 for each participating title, divided equally between HarperCollins and the authors. Wiley reported $49 million in AI-licensing revenue for fiscal year 2026 through strategic collaborations including Amazon Web Services, Microsoft, Anthropic, Perplexity, IQVIA and OpenEvidence.

Keith Riegert, president of the Stable Book Group and chief executive of Perfect Bound, calls himself a “hardcore techno-pessimist”. Yet he said he uses AI because “you use AI or you die at this point,” adding that its use must “start with a human and end with a human”.