A Dutch antiquarian bookseller named Pieter de Vries got an odd email this summer. A person named Natalia, working for a company called 2077AI, asked him to source over 3,000 books.
The email came with a spreadsheet listing more than 3,000 ISBNs. It asked him to match titles, prepare quotes and estimate shipping costs to China.
De Vries barely glanced at it. He figured it was spam or a phishing attempt and moved on.
Weeks later, a journalist investigating the request reached out to him. That's when he learned the order might be linked to artificial intelligence.
"I was shocked!" de Vries told Fortune. He said the request was "very unusual" for his business, which deals in rare and expensive books sold one at a time.
He wasn't alone. Several other Dutch booksellers received the same request and also dismissed it as a scam.
Why AI Companies Want Physical Books
The books listed in the spreadsheet were mostly published between 2020 and 2021. They came from academic publishers including Emerald Publishing, Elsevier, Wiley, Routledge and Oxford University Press.
The topics ranged from business and education to engineering, medicine and public policy. The email never mentioned AI or explained what the books were for.
This request surfaced as AI companies search beyond the open internet for quality written material. They need this text to train large language models.
Last summer, court records showed Anthropic had purchased millions of physical books for this purpose. The company removed the bindings, scanned the pages and threw away the originals.
This process is called destructive scanning. It involves cutting a book's spine so pages can run through high speed scanners before the copy gets discarded.
Anthropic's internal name for this effort was Project Panama, according to The Washington Post. The case later settled after a judge ruled that training AI on legally purchased books counts as fair use under copyright law.
Separate claims involving Anthropic downloading books from online libraries LibGen and PiLiMi were also settled. An Anthropic spokesperson said Claude models train on public web data, purchased datasets and internally generated data.
The company said it buys books through normal commercial markets. It said none of its data programs destroy rare or antiquarian books.
The Hidden Book Supply Chain
The court case revealed what happens after books are bought. It didn't explain how companies source millions of books in the first place.
Earlier this year, 404 Media reported that ISBNdb had begun advertising bulk book buying services for AI labs. The company's site reportedly offered purchases ranging from 1,000 to 1 million books "delivered at the scale AI demands."
Those webpages have since been removed. ISBNdb told Fortune the service never actually launched and it never bought or scanned books for any AI company.
Similar reports surfaced in Germany and Switzerland. Local media said secondhand booksellers there also received unusual bulk orders for specialized books.
A Canadian company named in some of these reports, Zoom Books, denied the allegations. It told Swiss broadcaster SRF the purchases were part of its regular recycling and trading business.
Fortune reached out to 2077AI for comment about the procurement request and whether the books were meant for AI training. The company had not responded by publication time.
