For years, people feared artificial intelligence (AI) would replace writers. Now critics say it is doing something even more unsettling: Consuming the books those writers created.
The latest controversy centres on Anthropic, one of the world’s leading artificial intelligence companies and the developer of the chatbot Claude. Like rivals behind ChatGPT and Gemini, Anthropic needs vast amounts of human-written material to train its AI models. Allegations surrounding one of its internal projects have now raised an uncomfortable question: How far should AI companies go in their race for better data?
According to ScreenRant, unsealed court documents from Bartz v Anthropic PBC reveal details of an internal initiative called Project Panama, which reportedly began in early 2024. The documents describe it as an effort to “destructively scan all the books in the world” while instructing employees not to discuss the project publicly.
Unlike ordinary book digitisation, which preserves the original copy, destructive book-scanning permanently destroys the book. The reported process involves buying books – including rare and out-of-print editions – cutting off their spines with hydraulic machines, scanning every page at high speed and pulping what remains. The digital copy is then used to train AI models.
The destruction is not technically necessary. Libraries, museums, and archives routinely use specialised scanners that preserve books while creating digital copies. Those methods, however, are slower, require more manual work, and cost considerably more. Removing the spine allows pages to pass rapidly through high-speed scanners, making mass digitisation faster and cheaper. Critics argue that commercial efficiency does not justify destroying physical books that could otherwise have been preserved.
Anthropic reportedly chose older books because they predate generative AI and are entirely written by humans. Researchers have warned that AI models trained heavily on AI-generated material become less reliable over time, making authentic human writing increasingly valuable. Critics counter that this still does not justify destroying culturally significant works when preservation-friendly alternatives exist.
For authors, librarians, and historians, the issue extends beyond technology. Rare books are historical artefacts, and many surviving copies cannot be replaced once destroyed.
The controversy is linked to Bartz v Anthropic PBC, a lawsuit filed by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, who alleged Anthropic used their books without permission to train large language models. The Authors Guild said the case produced a proposed $1.5 billion settlement covering roughly 500,000 books allegedly obtained from piracy sites, including LibGen and PiLiMi.
The case may also have created a far-reaching legal precedent. The court reportedly concluded that if a company legally buys a book, digitises it, and destroys the original, the process can qualify as “transformative” fair use. Under that interpretation, companies may train AI using those digital copies without the author’s permission.
The impact is already reaching beyond the courtroom. The Telegraph has reported a surge in bulk purchases of scarce second-hand books by buyers linked to AI companies, including titles with very few surviving copies.
The debate has now grown beyond copyright. It has become a broader question about who controls humanity’s written knowledge in the AI age and whether preserving books matters as much as teaching machines to read them. Unless a higher court overturns the current interpretation, companies such as Anthropic remain legally able to continue purchasing, digitising and destroying books for AI training under the existing legal framework.
