Summary: Reporting shows that major AI companies used third-party brokers to buy large numbers of printed books, scan them for training data, and then discard the originals. A 2025 court ruling found such mass scanning could be transformative and qualify as fair use; related litigation ended in a $1.5 billion settlement. Vendors such as ISBNdb briefly advertised large-scale, NDA-backed sourcing before removing the offer, while booksellers report dramatic spikes in rare-book orders, prompting ethical and preservation concerns.
AI Firms Secretly Buy — Then Destroy — Millions Of Books To Train Models
Credit: Getty Images
New reporting indicates that major artificial intelligence companies have been outsourcing the purchase of printed books to third-party brokers, digitizing those books for model training, and then discarding the physical copies. The practice—reported by 404 Media and discussed in documents unsealed during related litigation—has raised legal, ethical, and cultural-preservation concerns.
Credit: Cheng Xin via Getty Images
How It Worked
According to the reporting, AI developers sought out human-authored books published before the widespread use of AI-generated text because those texts are perceived as "cleaner" training material. To avoid direct association with mass purchases and the negative publicity that could follow, companies used intermediary services to source large volumes of titles from libraries, used bookstores, and out-of-print catalogs.
Credit: hapabapa via Getty Images
Key Legal And Corporate Details
In a high-profile 2025 case, a judge found that an AI company’s process of de-spining, scanning, and discarding millions of print books could be considered transformative and therefore might fall under fair use as described in Section 107 of the Copyright Act. The company later settled related litigation for $1.5 billion. Internal documents revealed during the case showed executives were aware that a legal defense might not prevent reputational harm.
Credit: FromSoftware
Internal note (reported): "We don't want it to be known that we are working on this."
Middlemen And NDAs
One intermediary named in the reporting, ISBNdb, briefly advertised a printed-book procurement service "tailored to your LLM training needs, delivered at the scale AI demands," and promised a "strict NDA on every engagement." The company removed the page after public scrutiny, saying the offer was exploratory and that it had decided to "pivot away" from that direction.
Market Effects And Bookseller Reports
Booksellers report a sudden increase in bulk orders for rare and low-circulation titles. One dealer told reporters he had seen orders jump from about 20 books a week to hundreds, particularly for volumes with ISBNs—suggesting buyers may be harvesting lists from ISBN-based databases. Similar trends were reported by some booksellers in the Netherlands earlier in the year.
Ethical And Cultural Concerns
Critics including librarians, booksellers, and cultural-preservation advocates argue that quietly pulping unique or hard-to-replace copies risks irreversible loss of cultural heritage. Even if courts accept scanning and disposal as legally defensible, the decision to permanently remove original texts raises questions about transparency, stewardship, and whether market forces should permit the consumption of physical culture as one-use training material.
What Comes Next
The collated reporting underscores a growing tension: developers’ demand for high-quality human-authored datasets versus public concerns about copyright, cultural preservation, and corporate transparency. The revelations may prompt calls for clearer industry standards, regulatory oversight, or voluntary commitments to protect irreplaceable physical works.
Help us improve.
