AI firms quietly buying rare books, destroying them to train models

The tension between AI training efficiency and book preservation reflects a deeper clash between speed-to-market and stewardship. A lawsuit in summer 2025 exposed Anthropic's destruction of millions of print books, and the backlash has forced the company to adopt a codename—"Project Panama"—for its destructive scanning efforts in an attempt to hide the practice from public scrutiny. The discovery has mobilized rare booksellers, who now recognize suspicious bulk orders as markers of AI involvement: large batches of unrelated titles purchased without price negotiation stand out sharply from typical library or university acquisitions. The industry's concern is not merely sentimental—it reflects a real threat to the survival of irreplaceable collections and an erosion of authors' intellectual property rights. Yet the economics of AI training create powerful incentives to cut corners. Non-destructive alternatives do exist: Google's 2009 patent, though imperfect, offers a blueprint, and the Internet Archive's labor-intensive hand-scanning—developed over decades and perfected by scanners like Eliza Zhang, who has processed more than 3 million pages and 18,000 items—demonstrates that preservation and digitization can coexist. The question now is whether public backlash and the commitment of firms like OpenAI and Microsoft to preserve public-domain collections will be enough to shift the broader industry toward preservation-focused methods, or whether cost pressures will continue to drive destruction.