Tobias Reithmeier

Blog

Project Panama: Why an AI Company Cut Apart Millions of Books

In January 2026, the Washington Post published an investigation that surprised even seasoned observers of the AI industry: starting in early 2024, Anthropic, the company behind the chatbot Claude, ran a covert operation called "Project Panama". It bought up millions of print books, mechanically sliced off their spines and scanned the loose pages at high speed. The paper then went straight to recycling. No physical archive was kept.

What the Court Records Show

The scale of the project became public not through the company itself, but through more than 4,000 pages of court documents unsealed in a class-action copyright lawsuit brought by authors against Anthropic. The records describe a genuinely industrial operation:

  • Anthropic spent tens of millions of dollars buying used books in bulk from libraries, online second-hand dealers and bookstores.
  • Service providers cut the books apart with hydraulic cutting machines and digitized the pages with high-speed scanners. "Destructive scanning" is far faster and cheaper than gentler methods that leave the book intact.
  • The program was led by an executive Anthropic hired in 2024 specifically for the job: the former head of book partnerships at Google Books.
  • An internal planning document from 2024 stated: "We don't want it to be known that we are working on this." The company knew exactly how the public would react.

The scanned texts were fed into the language models behind Claude as training data.

Fair Use, Piracy and 1.5 Billion Dollars

Legally, the case is more nuanced than the headlines suggest. In June 2025, US federal judge William Alsup ruled that buying print books, destructively scanning them and training on the result falls under the American fair use doctrine: the company had legally purchased the copies, and the court considered the training highly transformative.

What actually got Anthropic into trouble was something else. In parallel with the book purchases, the company had downloaded millions of books from piracy libraries such as LibGen, and that is precisely what the fair use ruling did not cover. In September 2025, a settlement of 1.5 billion dollars became public, roughly 3,000 dollars for each of the approximately 500,000 affected works. It is considered the largest publicly known payout in the history of US copyright law.

The irony is that cutting up books turned out not to be the legal problem in the end; reaching into the shadow libraries was. At the same time, the settlement shows how much value the industry places on high-quality text. A company that first spends millions on physical books and later pays billions in compensation has done the math and concluded it is worth it.

Traces Reaching German Used Bookstores

The phenomenon has long since arrived in Europe. As reported by the German public broadcaster Tagesschau, antiquarian bookstores in Germany and Switzerland noticed unusual bulk orders in 2026: a Canadian company was buying used books seemingly at random, in some cases up to 1,000 copies per dealer. The obvious suspicion is that here, too, stock is being acquired for AI training, because German-language literature is scarcer as training material than English-language text and therefore more sought after. The full supply chain has not been proven, and the company denies digitizing or destroying books itself.

Three Trends Project Panama Makes Visible

1. The open internet has been harvested. Wikipedia, forums, blogs and news sites are already baked into the models. Making language models better requires high-quality, edited, structured knowledge, and much of that sits in print books, specialist literature and archives. The industry's data hunger does not stop at the boundary between the analog and digital worlds.

2. "Move fast" now applies to copyright, too. Project Panama follows the familiar pattern of creating facts first and sorting out the legal questions later. Anthropic clearly priced in the litigation risk from the start. The fact that the bet partially paid off in court (fair use for purchased books) is more likely to encourage imitators than deter them.

3. The free-ride era of AI training is ending. The billion-dollar settlement points away from unpaid scraping and toward licensing and compensation models. Publishers, authors and rights holders are sitting at the table with real bargaining power for the first time. Humanity's collective knowledge is no longer a free raw material.

Conclusion

Project Panama is a bizarre anecdote about severed book spines, but it also shows that the race for the best AI has acquired a material dimension: warehouses full of books, hydraulic cutting machines, second-hand bookstores bought empty. The decisive battles of this era are fought in courtrooms as well as in data centers, and the courtrooms are where the value of humanity's printed heritage is currently being renegotiated.

Sources and Further Reading