AI firms are buying up rare books by the million, then destroying them
A hidden supply chain is funnelling centuries of printed knowledge into AI models, raising questions about cultural loss and copyright.
London’s antiquarian bookshops have become an unexpected front line in the race for artificial‑intelligence training data. Several dealers report receiving large, anonymous orders for obscure specialist titles, requests that mirror a pattern emerging across Europe. The buyers never identify themselves, decline to say where the books are going, and ask for extra photographs before committing to purchases that can run into thousands of volumes.
The trail leads to Anthropic, the San Francisco startup behind the Claude chatbot. Court filings reveal the company ran an internal effort codenamed Project Panama, buying millions of second‑hand books, slicing off their spines with industrial cutters, scanning every page and sending the remains for recycling. Executives considered books “essential” for teaching models “how to write well” rather than relying on what one co‑founder called “low‑quality internet speak”. Internal memos also show a desire to keep the operation out of public view.
A new supply chain takes shape
The ruling has accelerated a nascent market. Data‑broker ISBNdb now advertises bulk book‑buying services to AI labs, offering between 1,000 and one million volumes at a time. It pitches pre‑2022 publications as “structurally guaranteed” to be free of AI‑generated text, a selling point as synthetic content floods the web. The company’s own website acknowledges the reputational hazard: “‘AI company destroys two million books’ is not a headline that generates sympathy.”
Legal settlement and lingering questions
Anthropic separately agreed a $1.5 billion settlement over earlier use of pirated digital books, but the physical‑book programme continues under the fair‑use shield. Booksellers say the anonymous bulk orders have surged since the judgment, suggesting other labs are adopting the same playbook. For now, the trade operates in shadows, no names, no company letterheads, just pallets of centuries‑old knowledge disappearing into scanning facilities.
What happens next
Regulators in the UK and EU are watching. The UK Intellectual Property Office has signalled it may review whether mass digitisation of legally owned books for commercial AI training stretches copyright exceptions too far. Meanwhile, librarians and heritage groups warn that once a rare edition is pulped, its physical history, marginalia, binding, provenance, is gone forever. The next test case may decide whether the world’s libraries become protected archives or open quarries for the next generation of models.
