Project Panama: Anthropic Is Shredding Millions of Books to Train Claude

Picture a huge warehouse stacked with pallets of books. It is not a library or a quiet archive. Nobody reads here; they dissect. Workers in overalls carefully slice the spines off brand-new volumes, turning bindings into stacks of loose sheets that then vanish into high-speed scanners. This is not a scene from a Ray Bradbury dystopia but the daily routine of a secret unit at Anthropic. The project has the code name Project Panama, and its details have just surfaced in court archives, leaving the industry a little uneasy at the scale of what is happening.

Anthropic spent a long time building an image as the "ethical" alternative to OpenAI. While Sam Altman aggressively vacuumed up the internet, the makers of Claude talked about safety and responsibility. It turns out that when high-quality data runs short, ethics gives way to an industrial meat grinder. Court materials revealed that in early 2024 the startup's leadership launched an ambitious plan for the "destructive scanning of all the books in the world." The wording sounds ominous, but technically it makes sense: to scan quickly and without distortion, a book has to be literally destroyed and turned into a set of flat pages.

Why go to such trouble when e-book versions exist? The answer lies in quality and rights. Legal digital libraries are expensive and protected by strict licenses, while pirate archives such as Shadow Libraries often contain text recognition errors. Training models at the level of Claude 3.5 or the future Claude 4 requires clean, structured knowledge. Anthropic decided it was simpler and cheaper to buy up millions of physical copies, reduce them to scrap and get perfect digital versions than to negotiate with every rights holder individually. The operation's budget came to tens of millions of dollars, a sum comparable to the cost of buying H100 chips.

This situation highlights the central problem of the modern AI industry: the "data wall" is not a myth but a reality. Humanity has already fed neural networks nearly all of Reddit, Wikipedia and digitized newspaper archives. But the models' appetites grow exponentially. We used to talk about AI replacing writers; now we see AI literally devouring their physical legacy. The irony is that a startup valued in the billions has to deal in waste-paper logistics to gain an edge of a few percentage points in a chatbot's answer accuracy.

The secrecy around Project Panama has a simple explanation: it looks terrible from a PR standpoint. It is hard to sell the public on "safe AI" built on the ruins of destroyed books. Anthropic's lawyers probably hoped that physically owning a book would give them a loophole in the "fair use" doctrine. The argument goes: we bought the book, we have the right to read it, even if the "reader" is an algorithm and the reading process requires destroying the medium. Courts, however, are unlikely to be so kind to mass industrial copying.

What does this mean for us? We have entered an era in which information in the physical world is becoming more valuable than digital dust. We used to digitize books to preserve them for posterity; now we do it to feed a "black box" that hands us a summary in a chat window. This is a radical shift in how we consume knowledge. Soon we may face a shortage of rare editions simply because some new AI unicorn decides to buy up an entire print run to train its latest "language machine."

The bottom line: Anthropic has shown that in the battle for data, no prisoners are taken. Are we ready for the intelligence of the future to be built on the ashes of burned books?