Anthropic Bought Millions of Books to Scan for AI. Then It Shredded the Originals.

Anthropic, the AI company behind the Claude chatbot, ran an internal program called "Project Panama" that purchased millions of printed books, removed their bindings, scanned each page into digital training datasets, and shredded the physical copies. The practice is described in court filings unsealed in January in Bartz v. Anthropic, a class-action copyright suit brought by authors and first reported by Epoch Times.

The unsealed filings name two of the major used-book retailers from which Anthropic acquired inventory: Better World Books and World of Books. Anthropic's acquisition team was led by Tom Turvey, the former head of partnerships for Google's own book-scanning project, who emailed major book distributors and retailers about bulk purchases. The internal code name "Project Panama" appears in a 2025 court document in the case.

In the secondary book market, the program is a new kind of buyer: one that turns printed inventory into nothing. Books that flow into wholesale channels cannot be tagged at the seller level, and smaller retailers interviewed in the reporting describe no easy off-ramp once stock leaves their shelves. Charlie Becker, who has run a family bookstore in Houston for three decades, told Epoch Times the practice is an unprecedented challenge to bookselling and book preservation.

Anthropic, asked about the program by Tom's Guide, did not address the scan-and-shred pipeline directly. A company spokesman said: "Sourcing books is a widely used approach for training large language models across the AI industry" and "None of our data acquisition programs buy and destroy rare or antiquarian books." The statement leaves open the central question: whether Anthropic systematically destroys the books it buys.

Anthropic's denial of rare-book destruction is narrow. It does not contest the broader practice of buying in bulk, scanning, and discarding. That matters because the second-hand book market is not structured to distinguish rare from common inventory at the wholesale level, and because the volume Anthropic is alleged to have purchased, millions of books, far exceeds what any rare-book channel could supply.

What the books become after a wholesale channel fails to resell them is, in the normal course, pulp. That is the strongest counterargument: Anthropic's pipeline is just accelerated pulp. The argument strains in light of the structure described in the filing. The operation was led by a former Google book-scanning head, staffed to email major retailers about bulk purchases, given an internal code name, and run at a scale that cleared inventory at named wholesalers. The shape of the program looks more like a procurement pipeline than the residue of a remainder bin.

Bartz v. Anthropic is the only AI-book class action in which a buyer is alleged to have physically destroyed the source material. Authors Alliance, a writers' advocacy group, maintains a running tracker of AI class-action cases on the book side through early 2026, and Bartz is the standout. That makes the case a test of whether copyright law treats training-data procurement as a fair-use question, a destruction-of-source question, or both.

Booksellers do not currently have a way to mark inventory as "not for AI training buyers." Wholesale channels are not required to disclose the downstream buyer's identity or use. And the destruction of the physical book, once scanning is complete, means the original artifact is gone regardless of how a court eventually rules on the digital copy. Anthropic's answer, when it comes, will most likely arrive in court filings, not in a change to the procurement pipeline.