Project Panama: the secret to training AI
Anthropic’s internal memo outlines an initiative called Project Panama as the company’s effort to 'destructively scan all the books in the world.' This project was designed to enhance the capabilities of its AI language model, Claude, by feeding it a vast dataset of high-quality language from books.
According to the court documents, Anthropic employees were instructed to keep the project discreet, with the memo stating, 'we don’t want it to be known that we are working on this.' This level of secrecy suggests a mix of legal caution and sensitivity about public perception of the company's methods.
Anthropic had two main options for acquiring the data needed to train Claude. One was to legally obtain permission from copyright holders, but this was described internally as a 'legal/practice/business slog'—a time-consuming and complicated process. Initially, the company bypassed copyright laws by using pirated sources, but as that approach became too problematic, it shifted to destructive scanning as a more legally defensible solution.
Destructive scanning: process and purpose
The process of destructive scanning involves cutting the spines and edges of books to facilitate scanning. Each page is then trimmed to the right size and digitized before the original printed copy is discarded. This method, though effective for digitization, also results in the physical destruction of the source material.
Under the 'fair use' doctrine in U.S. copyright law, transformative use of copyrighted works is permitted without permission. Anthropic argued that scanning and digitizing books was transformative because the physical originals were destroyed, leaving only a digital copy for the company's use in training its AI. They framed the act as akin to teaching a human to read in order to write effectively.
The court found no evidence that the digitized books were distributed, sold, or made public. The destruction of the physical copies was part of a calculated strategy to minimize legal risks and streamline the logistics of data acquisition for AI training. Photos provided in court documents show warehouses filled with neatly stacked and labeled books, waiting to be processed.
The implications of destructive scanning
The U.S. lacks strong legal protections for printed books, and there are no endangered lists for printed works or clear regulations on their preservation. This lack of legal oversight raises concerns that other AI companies might adopt Anthropic’s methods, leading to the large-scale digitization and disposal of printed books. The result could be the loss of rare or unique texts that were never intended for mass reproduction.
While the court did not find copyright infringement in Anthropic’s actions, the destruction of physical books adds a significant ethical dimension to the case. Judge William Alsup noted that the company’s process was not about sharing or profiting from the content, but about extracting value through AI training. Yet the act of physically erasing books raises questions about the cultural and historical value of the printed word in a world increasingly dominated by digital reproduction.
As AI companies continue to seek large, high-quality datasets, the use of destructive scanning and similar methods could become more widespread. The precedent set by Bartz v Anthropic PBC may encourage others to pursue similar strategies, potentially reshaping the future of book preservation, author rights, and the balance between innovation and tradition in the digital age.
This case also highlights a broader issue: the evolving relationship between AI training and intellectual property. While training an AI on digitized books may be legally permissible under U.S. law, the implications for book culture, authorship, and the physical legacy of literature remain uncertain. The question of how many warehouses filled with destructively scanned books would be too many is one that lawmakers and industry leaders may soon need to address.