Correction Record / Models The AI Books Story Everyone Shared Is a Sales Page A used-book broker is telling AI labs that buying and shredding print books is now a court-approved way to get training data. That pitch is real and its legal skeleton holds up. The page selling it misstates the ruling it is built on and quotes a federal judge saying something he never said. Last updated: 26 July 2026 On 21 July 2026, 404 Media reported that ISBNdb, a book-metadata company, is brokering bulk print-book purchases for AI companies. The story travelled fast. Slashdot, Boing Boing, The Next Web and a run of newsletters and social accounts carried it within 48 hours, most working from the two paragraphs 404 Media leaves outside its paywall. The underlying business is real and worth reporting. What got added on the way out is a market that has not been shown to exist, a timeline that is four months off, and a legal account that came from the seller and was never checked against the court file. Disclosure: this piece turns on a ruling against Anthropic, and Anthropic makes the model used to draft this publication. What does the court file actually say? Judge William Alsup issued his order on fair use in Bartz v. Anthropic on 23 June 2025, in the Northern District of California, docket 3:24-cv-05417. It runs 32 pages and splits three ways. Training large language models on books was fair use, and the order calls that use exceptionally transformative. Converting purchased print books into digital files was also fair use, for a narrower reason: Anthropic stripped the bindings, scanned the pages, and discarded the paper, so one copy replaced another rather than adding one. Downloading over seven million pirated books to build a permanent library was not fair use, on every factor. The order awards nothing. It closes by setting a trial on the pirated copies and the damages arising from them, and notes that buying a book later does not absolve the earlier theft. That is the complete record on the point ISBNdb is selling against. Everything beyond it is the seller's. What the pitch added Each record sets what was published against what the primary source states. Read them and judge for yourself. Reported Headlines across Slashdot, Boing Boing, The Next Web and others say AI companies are buying old books in bulk. The framing treats a market as established. Record The evidence is one broker's marketing page. ISBNdb advertises a binding non-disclosure agreement on every engagement and names no client. No AI company is documented buying through it. The one company documented buying print books at scale is Anthropic, from court records, and it did so directly. Reported The story ran the week of 21 July 2026 as news. Record The ISBNdb post carrying the quoted lines is dated 24 March 2026 and was last updated 1 June 2026. It had been public for four months. Reported Labs are buying print because pre-2022 books are free of AI-generated text. Record ISBNdb's own book-sourcing page makes a different argument entirely. It says millions of valuable books were never digitized, and pitches low-resource languages and domain fine-tuning. It does not mention AI-generated text. The contamination framing sits elsewhere on the site. Reported ISBNdb tells prospective buyers the judgment was swift and expensive, at $1.5 billion in damages, with mandatory deletion and an injunction. Record The order awards no damages. It ends by scheduling a trial on the pirated copies and the damages arising from them. The $1.5 billion was a class settlement filed in September 2025 and approved this month. The deletion of pirated material was a settlement term, not a ruling. Reported ISBNdb prints a sentence attributed to Judge William Alsup, saying that a model built on stolen data becomes evidence of the theft . Record That sentence does not appear anywhere in the 32-page order. It also does not appear anywhere else on the searchable web except ISBNdb's own page. A second quotation on the same page, about a model containing what books knew, carries no attribution at all and is likewise absent from the order. The story that is actually supported Strip the additions and something more useful remains. A supply chain is forming around a court-blessed method, and its terms are visible on a public page. ISBNdb's sourcing page offers orders from 1,000 units up to 1 million titles, filtering by subject, genre, publication year, language or author, worldwide shipping to scanning facilities, and delivery structured for ingestion. Its blog adds a binding non-disclosure agreement on every engagement and documentation built to support a fair use claim: invoices, lot records, chain-of-custody material. It says plainly that the optics are a problem, and that a company destroying two million books is not a headline that attracts sympathy. That is the news. Not that labs are buying, which nobody has shown, but that the paperwork trail a lab would need to defend the practice is now a product someone sells, four months after the ruling that made it worth selling. The order itself supports why: Anthropic's defense rested on having purchased the physical copies and destroyed them, and a receipt is what proves the first half. Does this apply to you? If you write or publish books, the used-book market now contains a buyer category that wants volume and does not care about condition, and a water-damaged paperback serves the same purpose as a first edition. Authors receive nothing from a used sale, which is the point ISBNdb makes in its own ethics section. Ingram, the largest US book distributor, has offered publishers an opt-out. If you repeat things for a living, the operative lesson is narrower. Five outlets carried a page that misdescribes a ruling anyone can download, including a quotation from a sitting federal judge that a single search would not confirm. The ruling is free. Nobody opened it. Sources Order on Fair Use, Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA, Dkt. 231, N.D. Cal., 23 June 2025. Full text, 32 pages ISBNdb, print book sourcing for AI training, undated sales page, retrieved 26 July 2026. isbndb.com ISBNdb blog, on print book sourcing and the Anthropic ruling, 24 March 2026, updated 1 June 2026. isbndb.com/blog Emanuel Maiberg, 404 Media, 21 July 2026. 404media.co . Paywalled after the second paragraph; this piece relies on the free portion and on the primary sources above. Derivations None. No figure in this article is calculated. Every number is stated by the order or by ISBNdb's own pages, and each is linked at the point of use. Open questions Whether any AI company has bought books through ISBNdb. The company's stated non-disclosure terms mean it is unlikely to say, and no buyer has been identified. Whether ISBNdb stands by the quotation attributed to Judge Alsup, and where it came from. The company's account of the ruling has not been explained. What 404 Media's full article says beyond its opening. It sits behind a paywall, and this piece does not characterize the reporting inside it. Corrections policy AI Race Facts corrects visibly. Errors are fixed here with a dated note, never silently. This article quotes the court order directly and links it so any reader can check the two claims it disputes. Last updated: 26 July 2026. Related reporting: what the OpenAI escape coverage got wrong and the OpenAI detection gap .