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

In summer 2025, a lawsuit publicly revealed that Anthropic had destroyed millions of print books as part of its AI model training process. The disclosure sparked immediate backlash from book lovers and the broader publishing community, who fear that AI companies' hunger for high-quality long-form text will consume rare and irreplaceable collections. The core problem is straightforward: the cheapest and fastest way to digitize books is to cut off their spines, feed pages into scanning machines, and discard the originals. This method allows AI firms to process millions of titles quickly—a critical advantage in the race to advance their models—but it means rare books are lost forever once destroyed.

In response to the outcry, alternative technologies have been highlighted. Google patented a non-destructive scanning method in 2009, but studies have found that page curvature can distort text and pages can be missed, and workers may move too quickly, introducing errors like hands obscuring pages. The Internet Archive has taken a different path entirely, choosing labor over speed. A 2021 blog post detailed the work of Eliza Zhang, a book scanner at the Archive since 2010, who hand-scans rare books one page at a time. She raises the scanner glass with a foot pedal, adjusts cameras, and ensures readability before moving to the next page. She also documents fold-outs and inserts to prevent losing bonus materials. Her proprietary software halts the process if pages are skipped or images blur, prompting rescans. Zhang had scanned more than 3 million pages, 14,000 foldouts, and 18,000 items (mostly books) with a goal of zero errors. The Internet Archive rejected automated scanners with vacuum-powered page-turning arms because they damaged brittle books and rare volumes—the very collections librarians asked them to preserve.

Meanwhile, rare booksellers have begun flagging suspicious bulk orders. Last week, Kennys Bookshop in Ireland reported a "bananas" order for 5,000 obscure titles, with no attempt to negotiate price—a red flag that signals AI involvement. Unlike typical library or university purchases, these orders consist of widely varying, unrelated selections bought in massive batches at full price. Tomás Kenny of Kennys Bookshop told the Irish Times that AI is "terrifying our industry" and pledged to refuse orders suspected of destined for AI training to protect authors' rights and readers' access to hard-to-find titles. Booksellers worry that short-term revenue from such orders amounts to "signing their own death warrant," because once rare books are destroyed to train AI, they vanish from the supply chain forever.

Not all AI firms have taken a destructive path. OpenAI and Microsoft are partnering with Harvard librarians to train AI models on about 1 million public-domain books dating back to the 15th century. Elon Musk posted on X that he asked the xAI team to preserve rare books and scan them the hard way instead of cutting spines, though critics note this does not constitute a firm commitment to avoid destroying books entirely. Anthropic has denied destroying rare books and has provided no public accounting of its practices. The company reportedly warned its clients via ISBNdb—a book-database company advertising bulk sourcing services to AI firms—that the optics of book destruction were bad, and adopted the codename "Project Panama" to hide the effort from public view. Copyright law permits companies to destroy copies of books they purchase, but public backlash and the viability of preservation-focused methods suggest that the industry's long-term reputation may depend on choosing slower, more careful approaches over cost-cutting shortcuts.