Court Documents Reveal Anthropic's Secret Plan to 'Destructively Scan All the Books in the World'

Key points

• A federal court case in northern California, Bartz v Anthropic PBC, exposed an internal Anthropic project called "Project Panama" in late July 2026.
• An internal memo described the project's goal as "destructively scan all the books in the world" to gather training data for Claude, Anthropic's AI assistant.
• The memo warned staff to use the codename because Anthropic did not "want it to be known" that the work was happening.
• The project surfaces as AI companies face growing legal pressure over whether using copyrighted books to train AI systems requires author permission.

Anthropic builds Claude, a large language model, meaning software trained on vast amounts of text that learns to hold conversations, answer questions and generate writing. Like all such systems, Claude needs enormous quantities of written material. Where that material comes from is now the subject of lawsuits across the United States.

Court exhibit 21 in Bartz v Anthropic, filed in the Northern District of California, contains an internal memo that asks, bluntly: "What is Project Panama?" The answer written inside it: "Project Panama is our effort to destructively scan all the books in the world."

"Destructively" is a technical term from the book-scanning world: cutting the spine off a physical book so its pages feed through a scanner quickly. The book is ruined. It's faster and cheaper than flatbed scanning, but it's still scanning a copyrighted object without the author's consent.

The memo also explains why a codename was needed. "[B]ecause we don't want it to be known that we are working on this," it reads. That line now sits in a public court filing. Anthropic has not commented publicly on the memo.

Why does this matter to authors and readers?

Authors whose books were scanned get no say and no payment. Copyright law gives creators the exclusive right to reproduce their work. Whether feeding a scanned book into an AI training pipeline counts as that kind of reproduction is precisely what courts are working through.

The Bartz case, first reported by The Guardian, is one of several active suits testing this question. As we reported on 21 July 2026, Anthropic has already agreed to pay authors $3,000 per book in a separate $1.5 billion copyright settlement, approved by a federal judge, suggesting the company knows its exposure is real. A ruling against Anthropic in Bartz could force it to license books already used or pay additional damages.

For readers and Claude users, the immediate practical effect is small: the service keeps running while litigation proceeds. But if courts consistently side with authors, training costs could rise sharply, which would determine which companies can afford to build these systems at all.

What happens next?

No final judgment on liability has been handed down. Authors who believe their work was used without consent can check whether their titles appear in datasets that researchers have made public, though no complete list of Anthropic's training sources exists yet.

Congress has held hearings, and several proposed bills would require AI companies to disclose what copyrighted material they trained on, though none have passed.

Common questions

Does Anthropic admitting this mean it broke the law?

Not automatically. Using copyrighted material for AI training may qualify as "fair use", a legal exception that permits limited reproduction without permission. Courts are still deciding whether that defence holds here.

Could this affect other AI companies?

Yes. Similar lawsuits target OpenAI and Meta over training data. A ruling in Bartz would set a precedent that every major AI developer is watching closely.