Unsealed court filings reveal the name of a secret initiative aimed at destroying millions of print books to train AI — and how one of the leading companies in the field used it as a clever way to sidestep the copyright law.
AI startup Anthropic was involved in a secret initiative, known as Project Panama, to scan and dispose of up to two million physical books to train its chatbot Claude. “Project Panama is our effort to destructively scan all the books in the world,” the unsealed internal documents state, as quoted by The Washington Post. “We don’t want it to be known that we are working on this.”
But did they really want to keep it a secret? “There is no such thing as bad publicity,” it has been said. Indeed, the competition to dominate the AI market is fierce, with ChatGPT, Google Gemini, and Perplexity breathing down Claude’s neck. That could explain why Anthropic went to such lengths to speed up the training of its premium chatbot.
Love reading and writing? Submit a post to our blog.
According to documents, the company used industrial hydraulic cutters to “destructively scan” volumes purchased in bulk from major retailers. Books’ spines were removed and pages were scanned at high speed before being recycled or pulped.
Anthropic hired Tom Turvey, a former Google executive who previously led partnerships for the Google Books scanning project, in February 2024 to serve as a company vice president. Turvey was tasked with the massive undertaking to obtain the books for AI training while avoiding the “business slog” of traditional licensing negotiations with publishers. Leveraging his experience with large-scale digitization, he came up with a copyright law workaround that utilized the first-sale doctrine.
The first-sale doctrine is a legal principle in U.S. copyright law that allows the owner of a legally purchased work to treat that specific copy however they wish, including reselling or destroying it, without the copyright holder’s permission. By spending tens of millions of dollars to purchase physical books in bulk, Anthropic invoked its rights under the first-sale doctrine to scan and destroy them.
District Judge William Alsup ultimately ruled this practice constitutes fair use, finding that because Anthropic had legally purchased the books and kept the resulting files for internal training rather than public distribution, the process was “transformative.”
A process is considered transformative when it uses copyrighted material to create something with a fundamentally different purpose rather than replicating the original work. In this case, the judge determined that the AI was not being trained to redistribute the books, but rather to “turn a hard corner” by using the data to learn how to reason and generate entirely new content.
I’m a freelance writer with 6 years of experience in SEO blogging and article publishing. While you’re here, get the latest updates by subscribing to my newsletter.
