Developers of large language models are acquiring vast quantities of printed books to improve the quality of their training data. The practice, first uncovered in London, is prompting rare‑book dealers in Canada and across Europe to confront new supply‑chain and legal challenges.
A London‑based antiquarian seller told a local newspaper that it had received two unusually large orders from anonymous e‑mail addresses. Each request listed specialist titles and asked for additional photographs. When the seller sought to confirm the buyers’ identities, the enquiries were ignored. Similar bulk‑purchase solicitations have been reported elsewhere in Europe, with booksellers saying they are approached for thousands of obscure works by concealed purchasers.
Industry monitoring by 404 Media earlier this week highlighted a sharp increase in such transactions among second‑hand dealers.
Internal court filings show that Anthropic, the creator of the Claude model, launched an internal programme called Project Panama. The project classified physical books as “essential” for teaching AI systems to write well, preferring them to the low‑quality language found online. To assemble the dataset, Anthropic bought millions of second‑hand volumes, stripped the spines with industrial cutters, scanned every page and then recycled the remaining material.
A U.S. federal judge later ruled that the scanning operation qualified as fair use because the books were lawfully purchased, digitised and the originals destroyed rather than simply copied. Anthropic has previously agreed to a $1.5 billion settlement over allegations linked to pirated digital books.
The emerging demand has created a new market for printed works. Suppliers now offer orders ranging from one thousand to one million books per request to AI laboratories. ISBNdb, a broker that arranges bulk purchases, describes pre‑2022 printed books as "the world's best AI training data" and requires strict non‑disclosure agreements to protect buyer anonymity.
While the influx of high‑quality text promises richer AI outputs, it also carries reputational risk. ISBNdb admits that headlines about AI firms destroying millions of books do not generate public sympathy.