Booksellers report that artificial intelligence companies are bulk purchasing rare books, then removing them from circulation. The practice raises concerns about content acquisition methods used to train AI models.
Rare book dealers across multiple markets have noticed a pattern of coordinated purchasing of obscure and out-of-print titles, followed by the books disappearing from resale channels. Booksellers suspect AI training firms are acquiring these volumes to build training datasets, then destroying physical copies to prevent duplication.
The purchases typically involve single or small quantities of hard-to-find books at premium prices. Once acquired, the books do not reappear on secondary markets, suggesting they are being retained for data extraction rather than resale.
This buying pattern differs from typical publisher or collector behavior. Industry observers note that large-scale book procurement for AI training has become increasingly common, as firms seek diverse text sources to improve language models.
Booksellers have not publicly identified specific companies, though the practice aligns with known AI training methodologies that require extensive text corpora. The situation highlights growing tensions between AI development practices and traditional publishing and book preservation sectors.
Scientists at IIT Bombay and Adobe Research have developed a method to reconstruct original prompts from LLM outputs with near-perfect accuracy. The technique, called "Previous-Token Prediction," poses significant security risks for companies using proprietary system prompts.
Geoffrey Hinton, Fei-Fei Li, and Andrew Ng discussed AI regulation and open-source development at the Ai4 conference, weighing safety concerns against competitive pressures from China.
SpaceX's AI division has released Grok Bot, an always-on AI agent service that can independently complete workplace tasks by signing into your existing apps and tools.