A Dutch bookseller's 3,001-title order exposed a widening trade in which AI firms buy used books by the thousand, scan every page, and discard the originals — a practice a federal judge ruled legal even as Anthropic paid $1.5 billion for the pirated copies it used before it started buying print.
Used-book dealers are accustomed to strange requests — a single out-of-print title, a signed first edition, the occasional collector chasing a specific printing. What they are not accustomed to is an order for 3,001 books at once, with no theme except that they were mostly published between 2020 and 2021.
That is what landed in Pieter de Vries’s inbox in July. De Vries runs a second-hand bookshop in Haarlem, Netherlands, and the order came from a buyer identifying itself as “2077AI,” attached to a spreadsheet of ISBNs and instructions to ship to China, according to Startup Fortune. He nearly deleted it as spam. “I was shocked,” he told Fortune, once he understood weeks later what the list actually was: academic backlist from Emerald Publishing, Elsevier, Wiley, Routledge and Oxford University Press — engineering textbooks, public-policy monographs, medical titles nobody was clamoring to reprint. Nobody wants three thousand engineering textbooks for a shelf. Somebody wants three thousand engineering textbooks for a dataset.
the machine behind the order
Anthropic has been running exactly this kind of acquisition for two years, under an internal program called Project Panama. Court filings unsealed in Bartz v. Anthropic describe the company hiring Tom Turvey, the former Google Books executive, in February 2024 to buy physical books in bulk and run them through a hydraulic cutting machine that shears off the binding, feeding the loose pages into a high-speed scanner, according to DesignRush. One internal planning document quoted in the filings states the ambition without euphemism: “Project Panama is our effort to destructively scan all the books in the world.” Another, less plainly but more honestly: “we don’t want it to be known that we are pursuing this project.” An Anthropic spokesperson told Fortune the company trains on a mix of public web data, commercially acquired datasets and internally generated data, purchases books through regular commercial markets, and that none of its data-acquisition programs “acquire or destroy rare or antiquarian books” — a denial that sits uneasily beside the court record.
what the court settled, and didn’t
Judge William Alsup ruled in June 2025 that the buying-and-scanning operation is legal. Digitizing a lawfully purchased book, then discarding the paper, counts as fair use because “every purchased print copy was copied in order to save storage space and to enable searchability as a digital copy. The print original was destroyed. One replaced the other,” Alsup wrote, as Yahoo reported. Destroying the book, in other words, is what kept the practice legal — two copies would have been infringement, one is not. That ruling didn’t touch what Anthropic had done before it started buying: cofounder Ben Mann downloaded Books3, a 196,640-title pirated library, in early 2021, then five million more titles from Library Genesis and two million from the Pirate Library Mirror. For that, Anthropic agreed to a $1.5 billion settlement, approved July 20, 2026 by Judge Araceli Martinez-Olguin, covering roughly 482,000 works at about $3,000 each — the largest copyright recovery in U.S. history, per DesignRush. Federal judges have since issued similar fair-use rulings in separate suits against OpenAI and Meta, according to Yahoo Finance, extending the legal template industry-wide.
a market reshapes itself
The demand is spreading beyond one company. According to 404 Media’s reporting, cited across the coverage, suppliers have sourced anywhere from a thousand to a million books at a time for AI labs, specifically targeting titles published before 2022 on the theory that anything older predates the flood of AI-generated text now polluting the open web. ISBNdb, previously known for ISBN metadata, advertised a bulk-sourcing service “tailored to LLM training needs” and “delivered at the scale AI demands” before pulling the pages; the company later told Fortune the service was never launched and it never purchased or scanned books for AI clients, an account 404 Media’s reporting and DesignRush’s citation of an ISBNdb blog post — “The optics problem is real. ‘AI company destroys two million books’ is not a headline that generates sympathy” — sit in tension with. One unnamed bookseller told 404 Media weekly sales climbed from roughly 20 books to several hundred once AI buyers arrived: “It benefits me financially as well as by clearing out old inventory that is otherwise unlikely to sell… I don’t like that uncommon books are being pulped.” Canada’s Zoom Books, named in some reports, told Swiss broadcaster SRF the purchases were ordinary recycling and trading. In London, a dealer told City AM of two large anonymous orders requesting extra photographs before purchase. In Australia, Dawn Albinger of the antiquarian booksellers’ association says no confirmed bulk AI order has landed yet — “The tyranny of distance and the cost of shipping is one reason why we haven’t seen it,” she said — a reprieve of geography, not policy.
the question the ruling skipped
Fair use answers whether Anthropic had the right to train on the text; it says nothing about whether a rare or uncommon copy should be destroyed to get it. Elon Musk has said his SpaceXAI team will “preserve any rare books in a library and scan them the hard way vs just cutting off the spine and scanning.” Whether that alternative scales, and whether booksellers like de Vries or Albinger start demanding it as a condition of sale, is the thing to watch next — along with whether any archive steps in before the next 3,001-title order gets filled.
Project Panama is our effort to destructively scan all the books in the world.
