Some documents, dated 2024, reveal the existence of this initiative, described by the company itself as an effort to “destructively scan all the books in the world.” An image revealed in court documents shows a book warehouse that was allegedly part of Anthropic’s Project Panama, its project to scan, digitize, and destroy millions of books. An investigation by The Washington Post, based on court documents, reveals how the artificial intelligence company Anthropic developed an initiative called Project Panama to acquire, dismantle, and digitize millions of books in order to train its chatbot Claude.
The outlet claims to have examined more than 4,000 pages of court documents released as part of a copyright lawsuit against Anthropic. The documents, dated 2024, reveal the existence of this initiative, described by the company itself as an effort to “destructively scan all the books in the world.” The same document added an instruction that highlights the confidential nature of the operation: “We don’t want it to be known that we are working on this.” The court documents do not explain why the company chose the name Panama for the project.
They do detail that the initiative aimed to create a massive digital library to feed Anthropic’s artificial intelligence models. According to the documentation, the company spent tens of millions of dollars in approximately one year acquiring physical books. The procedure involved removing the spines of the books using industrial machinery, scanning the pages, and then recycling the materials. The information came to light after a federal judge authorized the release of numerous documents related to a lawsuit filed by writers accusing Anthropic of copyright infringement.
Photograph of a person holding a phone with the logo of the artificial intelligence (AI) company Anthropic, in New York (United States).
The company agreed to pay $1.5 billion to settle the litigation, although without admitting liability. According to an investigation by The Washington Post , in January 2023, one of the co-founders of Anthropic stated in an internal document that training models with books could teach them to “write well,” in contrast to the lower quality language found on the internet. Internal emails from Meta revealed in court proceedings described access to large digital collections of books as “essential” to compete in the AI market.
However, companies faced an obstacle: obtaining licenses directly from authors and publishers was complex and expensive. According to the lawsuits, several companies chose to obtain books on a large scale without the authorization of the copyright holders. Among the practices questioned were downloads from pirated digital libraries such as LibGen and Pirate Library Mirror. The documents reveal that Ben Mann, co-founder of Anthropic, personally downloaded an extensive collection of books from LibGen over 11 days in June 2021.
Benjamin Mann, Co-Founder of Anthropic
A year later he shared with colleagues the launch of Pirate Library Mirror, a platform that openly claimed to violate copyright laws in numerous countries. Anthropic maintains that it never used those collections to train a complete revenue-generating business model and claims that Pirate Library Mirror was also not used to train a comprehensive system.
When the company decided to abandon its reliance on pirated digital libraries and create its own repository, it launched Project Panama. To lead the effort, he hired Tom Turvey, a Silicon Valley veteran who had been involved in the creation of Google Books, the book digitization project launched by Google two decades ago.
Tom Turvey, Vice President of Product Partnerships at Anthropic.
The options considered included bulk purchases from second-hand bookstores and possible agreements with US public libraries. Finally, Anthropic acquired millions of books through specialized distributors such as Better World Books and World of Books. Although court documents conceal the exact figures, a business proposal included in the file indicates that the company sought to digitize between 500,000 and two million books in just six months.
The procedure involved using a hydraulic cutting machine to carefully separate the pages of each copy. These were then processed using high-speed industrial scanners, and once the digitization was complete, the physical remains were removed by recycling companies. The project’s unveiling comes amid a wave of lawsuits against artificial intelligence companies filed by writers, artists, photographers, and media outlets.
Google, Meta, Microsoft, and OpenAI also face similar lawsuits for using copyrighted works to train AI systems. So far, several US federal judges have issued preliminary rulings in favor of technology companies regarding the training of models. In June, Judge William Alsup concluded that the use of books to train artificial intelligence could be considered a “transformative” use and therefore compatible with the fair use doctrine.
Alsup compared the process to the teaching a student receives when learning to write from existing works. However, the judge distinguished between the use of the books and how they were obtained. While he deemed the project to digitize physical books legal, he noted that Anthropic may have infringed copyright by previously downloading millions of pirated works.
An Industry Under Scrutiny
The documents also show that Meta faced similar internal debates. Company employees expressed concern about downloading millions of books via file-sharing networks. “Downloading torrents from a corporate computer doesn’t feel right,” an engineer wrote in 2023.
Another internal message indicated that the use of LibGen had been authorized after a consultation with “MZ”, an apparent reference to CEO Mark Zuckerberg. The communications also reflect concern about the potential regulatory impact if it became public knowledge that the company used pirated material to train its models.
