Old printed books are selling in bulk because they contain no artificial intelligence text. Companies need this specific data to train their models without the errors that result from feeding AI systems content generated by other AI systems.
Why pre-2022 books matter
ISBNdb describes this material as curated, peer-reviewed, and authoritative. The company states that web crawls cannot replicate this structure. Books published before 2022 are considered structurally clean of modern poisoning tools.
Using data generated by AI for training causes model collapse. This process degrades model performance and increases error rates. Authors can also sabotage models by writing text designed to manipulate the resulting algorithms. ISBNdb claims that physical books from the pre-large language model era are the only guarantee against this contamination.
How the business works
ISBNdb previously managed inventory for libraries and distributors. The generative AI boom has shifted its focus. The company now sources bulk printed book purchases ranging from 1,000 to 1 million books per order. It helps AI labs acquire, scan, and process these texts while avoiding duplication.
The legal landscape
Attention to this activity rose in January following a copyright lawsuit. Book authors sued Anthropic, revealing internal documents detailing a plan to obtain and scan millions of printed books. The plan included destroying the books after scanning. A Washington Post article noted that Anthropic bought books from Better World Books, a marketplace for used books.
Google faced a similar situation. Publishers recently sued Google Gemini for training on copyrighted books. Anthropic did not respond to requests for comment.
What it means
One bookseller told me he noticed a historic spike in sales starting in April. He suspects he sold hundreds of books to AI companies. He sells foreign language titles and rare editions. His inventory often includes books without ISBNs, which were not part of these bulk orders.
Normally, he sells about 20 books a week. Since April, he has regularly sold hundreds. He noted that library orders usually focus on specific subjects. Recent orders lacked a rhyme or reason. They included books with little in common except for having ISBNs. The seller also observed a disregard for price, with some books selling at greatly overpriced rates.
Mike Feldman, director of client services at Alibris, confirmed that new bulk buyers are scooping up trade books. A publication in the Netherlands reported similar large purchases in June. Rare booksellers assumed these orders came from tech firms.
It is difficult to prove the books are for training data because buyer identities remain hidden. Large orders go to distribution centers where quality is checked. Anthropic contracted Datamation to scan the books. Datamation offers high volume destructive and non-destructive book scanning services. In a destructive process, the spine is cut so pages can be fed into a scanning machine. This method is faster and cheaper.
A federal judge in the Anthropic lawsuit, William Alsup, ruled that creating digital copies was legal because the originals were destroyed. He wrote that the print original was replaced by the digital copy. There was no evidence the digital copy was shared or sold outside the company. The judge determined this was transformative and qualified as fair use under Section 107 of the Copyright Act.
ISBNdb cites this legal argument. It states that purchasing paper books at scale from the secondary market does not deprive creators of income. These books have already discharged their financial obligation to their creators. The company argues that recycling destroyed books is the completion of a book’s lifecycle.
