Full Breakdown
AI Firms Bulk-Buy and Destroy Physical Books to Train Language Models
8/1/2026, 5:13:33 AM
The Bulk-Purchase Scheme
Artificial-intelligence companies are acquiring millions of printed books, scanning them with high-speed industrial equipment, and then discarding the physical copies. Internal documents from Anthropic’s “Project Panama” describe a hydraulic cutter that removes spines so individual pages can be fed through scanners. The process, known as “destructive scanning,” is favored for its efficiency and lower cost. A federal judge ruled that using legally purchased books in this way qualifies as fair use.
Legal Context and Fair-Use Rulings
In a 2025 case, Judge William Alsup found the digitisation “clearly transformative,” stating that each purchased copy was copied to create a searchable digital version before the original was destroyed. Anthropic later settled a class-action lawsuit for $1.5 billion, acknowledging that its Claude models were trained on a mix of publicly available web data, commercially acquired datasets, and internally generated data.
Emerging Supply Chain: ISBNdb and Middlemen
Investigative reporting by 404 Media revealed that ISBNdb, which catalogs over 111 million titles, briefly advertised a service to source 1,000 to 1 million books per order for “LLM training needs.” The page has since been removed, and ISBNdb told reporters the service was never launched. Similar bulk-order spikes have been reported in the Netherlands, Germany, Switzerland, Canada, and the United States, with sellers noting that orders contain random titles identified only by ISBN numbers.
Impact on Booksellers
Dutch antiquarian Pieter de Vries received an email from “2077AI” requesting quotations for more than 3,000 titles published between 2020 and 2021, which he dismissed as spam. Around the same period, a professional bookseller reported a fivefold increase in weekly sales, moving from fewer than 20 books to several hundred after April. The orders appeared random, ignored normal pricing, and included many rare or out-of-print editions, raising concerns that once scanned, those copies would be gone forever.
Official Statements & Responses
An Anthropic spokesperson said Claude models are trained on a blend of publicly available web data, commercially acquired datasets, and internally generated data, and that the company does not acquire or destroy rare or antiquarian books. ISBNdb issued a statement asserting it has never purchased, scanned, or sold books for AI training, and that the removed landing page was merely a test of market interest.
Criticism & Opposition
Investor Michael Burry called the practice “evil incarnate,” arguing that destroying books for AI training is morally reprehensible. Venture-capitalist David Sacks criticized the approach as hypocritical, noting that Anthropic claims entitlement to train on all output while allegedly appropriating others’ work without compensation.
Conflicting Reports & Gaps
Anthropic’s public statements deny involvement with rare books, yet court filings and internal documents detail large-scale acquisition and destructive scanning of millions of physical volumes, including antiquarian titles. ISBNdb’s denial contrasts with archived screenshots of the now-removed service page that targeted AI developers. No independent audit has confirmed the exact number of books destroyed or identified which titles were most affected, leaving a gap in verification of the practice’s scale and selectivity.
Verbatim Quote
- “Here, every purchased print copy was copied in order to save storage space and to enable searchability as a digital copy.” — Judge William Alsup
