AI companies are increasingly purchasing large collections of books printed before 2022 to obtain high-quality, human-written training data that is free from AI-generated content, often referred to as "AI slop." As the internet becomes saturated with machine-generated text, researchers fear that training new AI models on synthetic content could lead to model collapse, where successive generations of AI gradually produce lower-quality outputs. Pre-2022 books offer a curated, edited, and authoritative source of information created before the generative AI boom.
A key player in this trend is ISBNdb, a company that sources physical books in bulk for AI developers. According to reports, the books are typically digitized by cutting off their spines and scanning the pages—a process that often destroys the original copies but allows for faster and cheaper digitization. ISBNdb also offers strict confidentiality agreements to clients, acknowledging that headlines about AI companies destroying millions of books could generate public criticism.
The demand for older books is driven by both quality and legal considerations. Unlike online content, pre-2022 books are unlikely to contain AI-generated text or deliberate data-poisoning techniques designed to disrupt AI training. Recent U.S. court rulings have also strengthened the legal case for training AI on purchased and scanned physical books, provided the copies are lawfully acquired, encouraging companies to expand these acquisition efforts.
The trend highlights a growing challenge for the AI industry: finding reliable, human-created data as the web becomes increasingly dominated by AI-generated content. While supporters argue that digitizing older books preserves valuable knowledge for future AI systems, critics worry about the destruction of physical books, copyright concerns, and the broader implications of using published works to fuel commercial AI development. As demand for high-quality training data grows, books have become one of the most valuable resources in the race to build more capable AI models.
