AI Companies Are Buying Tons of Old Books Because They're Free of AI Slop (www.404media.co)

AI companies are increasingly turning to old printed books as a valuable source of training data, as these texts are devoid of AI-generated content that could degrade model quality. ISBNdb, a leading book database provider, emphasizes that books published before 2022 represent curated, authoritative knowledge essential for robust AI training. The service allows companies to acquire large volumes of these pre-LLM books, ensuring they are structurally free from contamination that can lead to “model collapse.” This method of sourcing has gained traction, especially after revelations of AI firms like Anthropic planning to purchase and destructively scan vast quantities of books for training purposes. The significance of this trend lies in its dual implications for the AI/ML community and the literary world. For AI developers, using high-quality, controlled data sources mitigates the risks of bias and errors in models that rely on web-scraped materials. Conversely, the destruction of printed books, particularly rare or specialized titles, raises ethical concerns among booksellers and authors, who fear loss of cultural heritage and commercial viability. Notably, ISBNdb offers anonymity in transactions, which shields AI companies from public backlash over book destruction. As the demand for clean data rises, the debate intensifies around the balance between innovation in AI and preservation of literary works.