Project Panama: How Anthropic Secretly Destroyed Millions of Books to Train Its AI
The Hidden Cost of AI Training
All AI models require training. How would an image generator program know what “Ghibli-style” art looks like unless its developers feed it hundreds of images from Ghibli movies? Odds are, you have trained AI programs without knowing, but most models draw from existing content and just Frankenstein the information together without removing the originals.
Anthropic’s LLM is not like most models.
Recently unsealed court documents reveal how Anthropic secretly bought, destroyed, and scanned millions of books to train its AI chatbot, Claude. It sounds like a Ray Bradbury story come to life. Somewhere in America’s sprawl of warehouses, book spines were sliced by machines like cuts of salami. Scanners then processed their contents in bulk.
What Is Project Panama?
Project Panama is a digital piracy and destructive book scanning operation by Anthropic aimed at developing a dataset of published old and rare books. According to an internal planning document, Project Panama is described as Anthropic’s “effort to destructively scan all the books in the world”.
The project began in early 2024 as Anthropic executives sought books predating those available online that could teach its AI chatbot Claude “how to write well” instead of mimicking “low quality internet speak”.
| Aspect | Details |
|---|---|
| Project Name | Project Panama |
| Company | Anthropic (creator of Claude AI) |
| Start Date | Early 2024 |
| Goal | Destructively scan all books in the world |
| Methods | Digital piracy + destructive book scanning |
| Lead | Tom Turvey (former Google Books executive) |
Why Physical Books Instead of the Internet?
AI companies have turned to using physical books because digital sources to train LLMs caused two major problems: legal liability and data quality.
The Problem with AI-Generated Content
The internet is now full of pages written by AI, and training new models on that text won’t allow it to become any smarter. Internal planning documents suggest Anthropic executives believed the initiative would lead to higher quality communication and address AI model collapse – a kind of degradation that occurs when AI trains on text contaminated by its own output.
What Books Offer
| Benefit | Description |
|---|---|
| Human-authored text | Free from AI interference |
| Professional editing | High-quality, polished language |
| Long-form reasoning | Complex narratives and structured arguments |
| Pre-2022 content | Predates the generative AI era |
How Destructive Scanning Works
The term “destructive scanning” wasn’t coined by an outlet reporting on the subject. Tom Turvey, Anthropic’s VP, came up with it. It’s faster and cheaper than the alternative.
The Process
-
Bulk Purchasing: Anthropic purchased millions of used and rare books from online retailers such as Better World Books, World of Books, and Zoom Books.
-
Spine Removal: Contractors used hydraulic cutters to slice off the spines.
-
High-Speed Scanning: Pages were fed into high-speed scanners and digitized in bulk.
-
Paper Disposal: The paper became waste, which was shipped off to recycling plants to be pulped and repurposed into products like toilet paper and cardboard boxes.
Note: This method is far faster than traditional book scanning. One vendor contract aimed to convert 500,000 to two million books over a six-month period.
The Scale of Destruction
The numbers are staggering:
Key Finding: By October 2024, Anthropic had already scanned (and effectively shredded) millions of books.
Is Destructive Scanning Legal?
Yes, under US copyright law. But the legal landscape is complex.
The First Sale Doctrine
The First-Sale Doctrine of the Copyright Act is what allows the buyer of a physical book to destroy it. Under US copyright law, the “fair use” doctrine allows you to make “transformative” use of copyrighted works without the owner’s permission.
The Court Ruling
US District Judge William Alsup ruled in 2025 that Anthropic’s scanning of legally purchased books for AI training qualified as fair use under US copyright law. The ruling described the use as transformative.
Key Quote: “Training” a corporate product is equivalent to training any human, teaching how to read in order to learn how to write.
The $1.5 Billion Settlement
Despite this ruling, Anthropic agreed to pay $1.5 billion** to settle a class-action lawsuit over allegations that it used pirated books to train its AI models. Authors typically received **$3,000 per book from this settlement.
The Rare Books Controversy
One of the most contentious aspects of Project Panama is the allegation that Anthropic targeted rare and antiquarian books.
What the Documents Show
Internal documents mention prioritising “less common” books, indicating that the company sought titles that were less widely available. However, they do not clarify whether this referred to genuinely rare books or simply books that were less popular or less commonly stocked.
Anthropic’s Response
Anthropic has stated: “None of our data acquisition programs buy and destroy rare or antiquarian books”.
The Fact Check
A fact-check by Snopes found no conclusive evidence that Anthropic specifically targeted rare or irreplaceable books for destruction. However, the claim that AI companies are buying and scanning millions of books is mostly true.
Industry-Wide Implications
Mysterious Bulk Orders
Bookstores have recently seen an uptick in “mysterious bulk orders,” specifically for “uncommon books” that are “unlikely to sell”. Booksellers in Ireland, Australia, the Netherlands, Germany, and Spain have reported similar patterns.
Not Just Anthropic
The running theory is that AI companies (plural) are behind these purchases, which, if true, means that Anthropic isn’t the only company running its own Project Panama. 404 Media tracked a shipment of rare books from California to an Amazon-owned warehouse with a logo of a Tyrannosaurus rex preparing to eat an open book.
Elon Musk’s Response
Elon Musk publicly posted on X that he “asked the SpaceXAI team to preserve any rare books in a library and scan them the hard way,” seemingly solidifying which side of the aisle he stands.
What This Means for the Future
Preservation Concerns
Concerns have grown among antiquarian dealers because some orders include uncommon or out-of-print works. As one bookseller noted, losing one copy of a book held by many libraries presents a different issue from destroying an edition for which only one copy is known to survive.
The Ethical Dilemma
The case has drawn attention to a fundamental question: Should we destroy physical books to build better AI?
| Perspective | Argument |
|---|---|
| Pro-AI | Books are legally purchased and scanned for transformative use. |
| Pro-Preservation | Destroying physical books, especially rare ones, risks losing cultural heritage. |
| Middle Ground | Non-destructive scanning alternatives exist (e.g., Treventus ScanRobot). |
Comparison Table: Project Panama at a Glance
| Aspect | Details |
|---|---|
| Project Name | Project Panama |
| Company | Anthropic |
| AI Model | Claude |
| Goal | Destructively scan all books in the world |
| Method | Hydraulic cutters + high-speed scanners |
| Books Purchased | Millions |
| Cost | Tens of millions of dollars |
| Legal Status | Ruled fair use (US court) |
| Settlement | $1.5 billion (copyright lawsuit) |
| Rare Books | Unproven (Anthropic denies) |
Frequently Asked Questions
What is Project Panama?
Project Panama is Anthropic’s secret initiative to buy, destructively scan, and destroy millions of physical books to train its Claude AI model.
Is it legal to destroy books for AI training?
Yes. The First-Sale Doctrine allows the buyer of a physical book to destroy it, and a US court ruled Anthropic’s scanning qualified as fair use.
Did Anthropic destroy rare books?
Anthropic denies this, and fact-checkers have found no conclusive evidence that rare or irreplaceable books were specifically targeted.
How many books were destroyed?
Millions of books were purchased, with one six-month contract covering 500,000 to 2 million books.
Why didn’t Anthropic just use e-books?
Managing copyright for e-books was deemed too complicated. Physical books offered a legally simpler path, even if it meant destroying them.
What happens to the paper after scanning?
The paper is shipped off to recycling plants to be pulped and repurposed into products like toilet paper and cardboard boxes.
Quick Reference Summary
| Step | Action |
|---|---|
| 1 | Anthropic identified the need for high-quality, pre-AI training data. |
| 2 | The company launched Project Panama to “destructively scan all the books in the world.” |
| 3 | Millions of physical books were purchased from vendors like Better World Books. |
| 4 | Hydraulic cutters removed spines, and pages were fed into high-speed scanners. |
| 5 | Digital copies were used to train Claude AI. |
| 6 | Physical copies were recycled into paper products. |
| 7 | A US court ruled the practice fair use. |
| 8 | Anthropic settled a related copyright lawsuit for $1.5 billion. |
Final Thoughts
Project Panama represents one of the most controversial chapters in the AI arms race. On one hand, it’s a testament to the lengths companies will go to secure high-quality training data. On the other, it raises profound questions about the value we place on physical books—and whether the pursuit of better AI justifies their destruction.
The bottom line: While the practice may be legal, the ethical debate is far from settled. As more AI companies adopt similar methods, the tension between technological progress and cultural preservation will only intensify.
