Anthropic's Covert Book Destruction Program Revealed Through Legal Filings
Key Takeaways
- ▸Anthropic spent tens of millions acquiring millions of books through Project Panama, a covert program designed to destructively scan texts for LLM training data
- ▸Internal documents show the company deliberately concealed the initiative, revealing awareness that public perception would be negative
- ▸U.S. courts ruled the practice legal under fair use and first-sale doctrines, enabling AI companies to purchase and destroy antiquarian books without copyright liability
Summary
Internal documents unsealed in January legal filings exposed Anthropic's covert operation known as Project Panama, under which the company has spent tens of millions of dollars acquiring millions of antiquarian books to be destructively scanned for AI training data. The books are processed using hydraulic-powered cutting machines that slice off spines and individual pages, which are then scanned with industrial-grade equipment before being discarded. Internal planning documents reveal that Anthropic was aware of how negatively the project would be perceived, with one memo stating "We don't want it to be known that we are working on this."
Despite these efforts at secrecy, the practice has proven legally defensible. A judge ruled that Anthropic's use of physical books is protected as fair use under U.S. law, with the company relying on the first-sale doctrine—a legal principle that allows purchasers to do whatever they wish with items they buy, including destruction. This doctrine, traditionally meant to enable the second-hand book trade, has created a loophole that permits AI companies to legally destroy rare books after extracting their contents. A $1.5 billion class action settlement was approved against Anthropic in connection with the use of pirated literary works in Claude's training.
The practice has spawned a new market segment among booksellers who now explicitly cater to AI companies seeking bulk quantities of antiquarian books. While some booksellers appreciate the financial benefit and inventory clearance, others express ambivalence about the destruction of uncommon and culturally significant works.
- A $1.5 billion settlement was approved in a class action lawsuit over use of pirated works to train Claude
- A secondary market for antiquarian books has emerged catering specifically to AI companies, creating ethical tensions between financial incentives and cultural preservation
Editorial Opinion
While legally defensible under existing U.S. law, Anthropic's deliberate concealment of Project Panama exposes a troubling gap between corporate behavior and public values in AI development. The company's explicit effort to keep the initiative secret—revealed only through legal discovery—suggests internal acknowledgment that mass destruction of books, however economically rational, conflicts with societal expectations around cultural stewardship and transparency. The reliance on first-sale doctrine to destroy millions of texts for private corporate gain may be technically lawful, but it highlights how outdated legal frameworks struggle to address novel threats to human-authored cultural artifacts in the AI era.
