AI Firms Buy and Destroy Used Books to Feed Large Language Models
Court documents reveal 'Project Panama,' an initiative to destructively scan physical books for AI training data.
The race to refine large language models has moved from the digital web to the physical shelves of secondhand bookstores. AI companies are increasingly purchasing bulk quantities of used books to acquire high-quality, long-form text that is free from the "AI slop" generated by other models.
Court documents have revealed an internal initiative at Anthropic codenamed "Project Panama," which focused on the "destructive scanning" of books. This process involves removing the spines of physical volumes to allow for rapid digitization, after which the remains are recycled. An Anthropic spokesperson stated that the company's chatbot, Claude, is trained on a combination of commercially acquired datasets, publicly available web data, and internally generated data.
The Bookseller's Dilemma
Independent booksellers are reporting an unprecedented surge in suspicious bulk orders. These buyers often exhibit specific "red flags," including the purchase of massive quantities of random, obscure titles and a notable lack of price negotiation. In one instance, Kennys Bookshop in Ireland flagged a "bananas" order for 5,000 obscure titles.
For some sellers, these orders provide a significant financial windfall. However, others are distressed by the implications. Tomás Kenny, owner of Kennys Bookshop, expressed opposition to the practice, stating that if books are being scanned specifically to extract intellectual property, his shop does not want to be a part of that process.
Legal and Cultural Implications
While the practice is ethically contentious, it has found some legal protection in the United States. In a 2025 ruling, U.S. District Judge William Alsup determined that using books purchased in this manner to train AI was "exceedingly transformative." The court concluded that such use did not constitute a copyright violation, effectively categorizing the process as fair use.
Despite the legal victory for AI firms, the practice raises urgent concerns regarding the preservation of human cultural heritage. Because destructive scanning permanently eliminates the physical copy, rare or unique editions may be lost forever in the pursuit of training efficiency. This creates a tension between the immediate profit of the secondhand market and the long-term survival of physical literature.
The Path Forward
As AI companies continue to exhaust high-quality digital data, the demand for physical archives is expected to grow. While some firms have collaborated with libraries for public-domain works, the shift toward the secondhand market suggests a growing hunger for diverse, non-digitized content. The industry now watches to see if further legal challenges will emerge or if booksellers will establish collective standards to prevent the permanent destruction of rare texts.