Pieter de Vries, an antiquarian bookseller at De Vries & De Vries in Haarlem, Netherlands, received an email from a person calling herself "Natalia" (or "Nataly") at a Singapore-based company called 2077AI. The message included a spreadsheet of more than 3,000 ISBNs — mostly niche academic and obscure titles — and asked him to source them, prepare quotes, and estimate shipping costs to China . De Vries initially dismissed it as spam or phishing. Similar requests have surfaced from used-book dealers across the Netherlands, Germany, Switzerland, and Spain, all suspected to be funneling books to American AI labs
."
Unsealed court documents from the Bartz v. Anthropic copyright lawsuit revealed that Anthropic ran a covert operation called Project Panama beginning in early 2024 . The company spent tens of millions of dollars buying millions of physical books, mechanically slicing off their spines, scanning every page, and destroying the originals — all to build a private, high-quality training corpus for its Claude chatbot
.
An internal planning document, dated 2024 and unsealed in January 2026, stated: "Project Panama is our effort to destructively scan all the books in the world" . The goal was to convert up to 2 million books in six months
. More than 4,000 pages of court documents detail the operation
.
On June 23, 2025, U.S. District Judge William Alsup of the Northern District of California ruled in Bartz v. Anthropic that using legally acquired books to train AI models is "quintessentially transformative" and constitutes fair use under U.S. copyright law . The key principle: as long as the AI company obtains the copies legally (buys them), scanning and destroying them for training is permissible
. The judge did, however, deny summary judgment on the issue of pirated copies — finding that downloading from piracy sites was not fair use
.
In July 2026, a federal judge in San Francisco approved a $1.5 billion settlement in the Bartz v. Anthropic class action, which authors (including Andrea Bartz, Charles Graeber, and others) had brought against Anthropic two years earlier . The settlement followed the June 2025 fair use ruling — Anthropic agreed to pay after the court had already established that its core practice (scanning legally bought books) was legal, but the case also involved claims over pirated material
.
A thriving secondary market has emerged for bulk book procurement targeting LLM training. Major AI providers use third-party middleman services — including the ISBN database company ISBNdb — to place anonymous orders for hundreds of thousands of books, shielding the end buyer's identity . Booksellers across Europe report near-identical bulk purchase inquiries. The purchases are shipped to hubs (sometimes in China or other processing centers) where the destructive scanning takes place
.
Archivists, librarians, and rare-book dealers have raised alarms that the last remaining physical copies of rare and out-of-print academic texts are being permanently destroyed . Unlike non-destructive digitization (e.g., the Treventus ScanRobot, which digitizes fragile books at 2,500 pages per hour without harm), industrial shredding is cheaper at scale — but it means no physical copy survives once a title has been processed
. For obscure niche works with only a handful of known copies in existence, this represents an irreversible cultural and scholarly loss
.