AI Companies Buy Rare Books Only to Destroy Them

All AI models require training. How would an image-generating program know what “Ghibli-style” art looks like unless its developers provide it with hundreds of frames from Ghibli films? Chances are you’ve trained AI programs without knowing it, but most models take inspiration from existing content and simply piece together Frankenstein’s information without removing the originals. Anthropic’s LLM is not like most models.

Anthropic has purchased millions of books, many of which are rare (or at least difficult to obtain), for the purpose of a special “destructive analysis” process. Called “Project Panama” in Anthropic’s internal documents, the company does not shred books after use, as some media outlets have previously claimed, but cuts them as part of the digitization process. Although an Anthropic spokesperson (via Fortune) said the company does not purchase “rare or antique books” with the intent of scanning and destroying them, the discovered documents tell a very different story.

According to an internal memo from April 2024, the Panama Project exists solely to “destructively analyze all of the world’s books”, not just rare or expensive books; everyone on the planet. Also, the name “Panama Project” is not a silly code name. In the memo, Anthropic Vice President Tom Turvey said the codename was intended to obscure the project’s activities — he didn’t want anyone outside the company to know what it was doing. This really makes you want to reconsider using Anthropic Claude.

The term “destructive analysis” was not invented by a media outlet covering the subject. It was Tom Turvey who invented it. Normally, when you think about scanning books, your mind probably goes to placing the item in an outdated photocopier or scanner and pressing the “Scan” button, then moving to a new page and repeating the process. But it’s too slow for Anthropic. When 404 Media reported on the Panama Project, it delivered two key takeaways.

First, the outlet claimed that Anthropic was buying old books in bulk because they were free of AI influence. What better way to train a model to write content that looks less like your standard AI slop? More importantly, Anthropic embraced destructive analysis because it is “faster and cheaper” than the alternative. Simply cut the book along the spine, separate the pages and insert them into a scanner.

We would like to point out that many companies sell devices designed to scan books without tearing them. These include SMA Australia’s Robo Scan range of robotic book scanners and the Treventus ScanRobot. Such devices could make it possible to scan books purchased by Anthropic without damaging them, a major concern among critics, especially as the company targets hard-to-acquire books. The public would be less interested in using Blu-ray players to preserve their retro gaming libraries if the devices shredded their game discs in the process, and the same is true here.

According to Anthropic’s own internal memo, Project Panama had already digitized (and effectively shredded) millions of books as of October 2024. To make matters worse, bookstores have recently seen an increase in “mystery bulk orders,” particularly for “rare books” that are “unlikely to sell.” The common theory is that AI companies (plural) are behind these purchases, which, if true, means Anthropic isn’t the only company running its own Panama Project.

However, due to a general lack of information, there is no way of knowing how many books have already been sacrificed in the name of AI models, much less whether any copies were the last of their kind. Although no one knows the true scope of destructive analysis processes, not all AI companies support this procedure. Probably. Hopefully.

Elon Musk posted publicly on However, remember when he claimed he was good at video games, but was accused of paying other gamers to make him look good? Or was he caught orchestrating a disinformation campaign about the US elections? Don’t be surprised if SpaceXAI is caught destructively scanning books in the near future.