Anthropic’s “Project Panama” — a plan to buy, scan, and destroy roughly 2 million books for Claude training — has reached its final procedural stage. Court records and reporting show the final approval hearing took place in San Francisco on May 14, 2026. Before that, on September 5, 2025, Anthropic agreed to a $1.5 billion settlement with the plaintiffs, followed by preliminary court approval later that month.
The court drew a line between purchased books and pirated files
The case was brought by Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson on behalf of about 500,000 authors in the Northern District of California, with Judge William Alsup presiding. In June 2025, Alsup ruled that using legally purchased physical books for AI training qualified as fair use. Books obtained through shadow libraries such as LibGen did not receive that protection.
That distinction shaped the case. The ruling did not treat all training material as one category. Instead, it separated lawfully acquired print books from unauthorized digital copies, making source and method of acquisition central to the copyright analysis.
Internal documents described a “destructive scan” operation
Documents disclosed in the case described Project Panama as a plan for “destructive scanning” of books and included a note stating that the company did not want the public to know it was doing this. The reporting says a professional scanning vendor processed between 500,000 and 2 million books over about six months. The workflow was blunt: cut off the spines, scan pages with high-speed equipment, then send the paper waste away for recycling.
The books came from sources including The Strand, online secondhand marketplaces, and libraries. They were ordinary used books that could have remained in circulation, not rare archival items. Anthropic’s legal theory relied on the first-sale doctrine: once a physical copy was lawfully purchased, the buyer could dispose of it. Destroying the original, the company argued, also prevented recirculation and supported a claim of transformative use. The judge accepted that reasoning for the legally purchased books.
Google Books offers a useful comparison, but not the same result in practice
The report compares the case with Google Books. Google publicly launched its scanning effort in 2004 through partnerships with major research libraries. It kept the original books, gave libraries digital copies, and let the public discover works through indexing and snippet preview. Publishers had an opt-out path. Google also won on fair use in its own copyright fight.
Anthropic’s project looked different. The content moved into a private model, the physical books were destroyed, and the program was kept secret. There was no visible opt-out mechanism. Both cases ended up within a fair-use framework, yet the practical effect was not the same: one expanded search and access, while the other fed a commercial AI system and removed the originals from circulation.
The settlement closed the case, not the debate
Based on the reported settlement amount, compensation worked out to roughly $3,000 per book. The report describes it as the largest copyright settlement in U.S. history. Even with the litigation wrapped up, larger questions remain open: who decides what knowledge enters a model, how that decision can be examined, and whether any mechanism should exist to challenge those choices before they are locked inside a private AI product.
The hearing may have finished the court process. It did not settle the argument over training data, access to books, or the tradeoff between private model development and the continued circulation of written knowledge.