AI labs accused of destroying ultra‑rare books for training data

AI companies are increasingly destroying ultra‑rare printed books to create high‑quality training datasets, raising alarms among archivists, historians, and booksellers. Pre‑2022 books are now highly valued because they contain human‑written text untouched by modern AI systems, making them ideal for avoiding model collapse and data poisoning. Bulk suppliers like ISBNdb reportedly ship up to a million books anonymously to AI labs, where high‑speed scanners cut spines, separate pages, and destroy the originals. Some of the volumes entering these pipelines are extremely scarce, with only a handful of surviving copies after centuries of wars, fires, and decay. A recent US court ruling involving Anthropic found that scanning and destroying legally purchased books can qualify as fair use, further enabling this practice. Critics warn that without public awareness, this quiet erasure of irreplaceable cultural artifacts could become one of the most significant losses of the modern era.

Read the full story on TechRadar →