The court files from a lawsuit against the AI firm Anthropic have revealed a contentious internal project named “Project Panama” that involved the purchase of physical books on a massive scale, scanning them, and subsequently destroying them to train its Claude AI model on the resultant data.
This has been revealed through unredacted court files from a copyright infringement lawsuit filed in 2025 by authors against Anthropic. The court files show that the project was ambitious and secretive, as indicated in an internal memo that described the project as “destructively scanning all the books in the world.”
Anthropic initiated the project, named Project Panama, in early 2024 with a clear objective of teaching its AI models “how to write well” by training them on high-quality text from published books. The firm sourced the books on a massive scale, sometimes tens of thousands at a time, from leading resellers such as Better World Books and World of Books.
How “Project Panama” of Anthropic Scanned and Shredded the Written World?
Once the books were obtained, they underwent an irreversible process. The books were cut open with hydraulic cutters to remove the spines, which enabled the pages to be scanned using high-speed scanners. Once scanned, the pages were sent to recycling, with no physical copies remaining. This was likely done to avoid taking up storage space and potentially raising different legal issues.
Internal company documents show that the company aimed at an estimated 130 million unique book titles worldwide, although the actual processing of millions of books was done through vendor partnerships. Proposals by scanning vendors indicated the ability to process between 500,000 and 2 million books within a six-month period, although the figures are partially redacted in court documents.
What is most interesting about Project Panama, however, is the level of secrecy that was maintained. Communications within the company urged employees to keep the project under wraps, as if the company expected public outcry if it were discovered that an AI firm was purchasing and destroying books on a massive scale.
The company did not limit itself to purchasing books from bulk resellers. There are mentions of attempts to contact famous bookstores such as The Strand in New York City and even libraries. Meanwhile, at least one of the founders of Anthropic downloaded files from LibGen, a “shadow library” that is famous for hosting pirated ebooks.
How Physical Destruction Secured a Fair Use Win for AI Training
The lawsuit that brought these facts to light involved whether Anthropic’s use of copyrighted books to train AI was a case of copyright infringement. Authors claimed that the company was profiting from their work without permission or compensation.
But in 2025, a judge ruled in Anthropic’s favor, ruling that the destructive scanning was a case of fair use under copyright law. This was based on a number of considerations: that Anthropic had purchased the books, rather than pirating them, that the scanning was done in-house rather than being shared publicly, and that the physical books were destroyed after scanning rather than being resold or kept.
From the perspective of Anthropic, this was actually a more legally sound course of action than downloading books from pirate sites. By purchasing the books and destroying them after scanning, the company was able to leave a paper trail of legal acquisition and avoid the gray area of using pirated digital copies.
Project Panama and the Secret Infrastructure of AI Training
Project Panama provides a glimpse into how far AI companies will go to get their hands on training data for their models. With the increasing complexity of these models, the desire for quality text, the kind that is professionally edited and published in books, has become enormous.
The case also provides a glimpse into a dilemma that the AI industry faces. The industry believes that it needs access to humanity’s written knowledge to develop useful AI models. However, authors and publishers believe that their work is being used for the benefit of the industry without fair compensation. The fact that a judge declared the process to be fair use does not provide any clarity on the ethical implications of whether this should be considered acceptable.
For now, Anthropic’s Claude models have access to the literary knowledge that has been derived from these millions of destroyed books. Whether other companies are following a similar process is not clear, but the secrecy surrounding Project Panama indicates that Anthropic was aware of the implications.
The unredacted court documents are a reminder that beneath the polished screens of AI-powered chatbots exists a huge, sometimes contentious infrastructure of data gathering—one that in this instance literally devoured the books it was learning from.
