Recently, a revelation about AI giant Anthropic on social media platform X (formerly Twitter) has sparked widespread attention. The report claims that the company is purchasing millions of books in large quantities, digitizing them through scanning, and then immediately destroying them. This move has been likened to a plot in a science fiction novel "The End of the Rainbow" by Vernor Vinge, written 20 years ago, triggering intense public debate about the methods AI companies use to obtain training data.

image.png

The "Panama Project" Comes to Light: The Cost from Piracy to "Legitimization"

According to court documents revealed at the beginning of 2026, this operation, named "Project Panama," was secretly carried out internally by Anthropic. The core goal of the plan was to obtain "all books in the world." For this purpose, the company hired senior executives who had participated in the Google Books project and invested substantial funds to purchase large numbers of books from secondhand book retailers and physical bookstores.

The operational process is striking: employees transported the purchased physical books to warehouses, performed "destructive scanning" by cutting off the spines, converting them into high-resolution PDF files. Subsequently, the remaining paper remnants were directly sent to recycling companies for destruction. This practice has sparked ethical controversy, but legally, Anthropic argues that it complies with the "first sale doctrine" and "fair use": since they have legally purchased copies, and the scans are used only for internal training without external distribution, destroying the originals is to ensure that no illegal secondary circulation of copyright copies occurs.

1.5 Billion Dollar Legal Cost: A New Strategy in the Copyright Battle

The exposure of "Project Panama" was not accidental. Previously, Anthropic was sued by several authors for allegedly obtaining data from the pirated e-book website LibGen. Its CEO once admitted that negotiating individual authorizations with publishers was extremely challenging both commercially and practically.

To address legal risks, Anthropic paid approximately 150 million dollars in settlement fees in 2025 to resolve collective lawsuits against its early pirated data sets. Afterward, the company began to shift toward the "Panama Project," a more expensive but lower legal risk model of acquiring physical books. Currently, judges tend to believe that this "legally purchased + scanned training" model has a strong defense foundation.

Truth and Boundaries: The Exaggerated "Cultural Disaster"

Although the event has been described online as "distilling the human knowledge base" or "a disaster for ancient texts," the reality is not so extreme. According to investigations, the destroyed books mainly included common secondhand books with wide market circulation, not rare ancient books or cultural heritage. The groups most affected are writers and publishing associations who believe their copyrights have been damaged, rather than cultural institutions.

Currently, there is still no final conclusion in the U.S. legal community regarding whether AI training constitutes "fair use," and other major companies including Meta and OpenAI are also deeply involved in similar legal quagmires. Although Anthropic's approach of turning millions of physical books into "digital ashes" technically completes high-quality data iteration, the moral costs and legal boundaries behind it remain pressing issues in the AI era.