Anthropic Destroys Millions of Books for AI Training Data

AI developers are purchasing millions of physical secondhand and out-of-print books to extract high-quality, human-authored training data for large language models. This process, known as destructive scanning, involves using industrial hydraulic cutters to remove book bindings and high-speed scanners to digitize pages before the original paper is pulped or recycled. These companies target pre-2023 volumes to avoid model collapse caused by training on AI-generated content found on the internet.

Anthropic PBC has led this trend through Project Panama, an industrial-scale operation designed to train its Claude AI. The company's actions follow a ruling by Judge William Alsup in the Bartz v. Anthropic case, which determined that the one-for-one transfer from physical to digital copy constitutes transformative fair use. Despite this, Anthropic agreed to a $1.5 billion settlement on July 20 to resolve separate claims involving the use of 482,000 pirated books.

Intermediaries like ISBNdb and Zoom Books have facilitated these bulk acquisitions by sourcing volumes from libraries and estate sales. ISBNdb previously advised clients to frame the activity as digital preservation to avoid public backlash. While rare booksellers in the United States, Europe, and the Netherlands report a surge in suspicious orders, Elon Musk has publicly opposed the practice, instructing his SpaceXAI team to preserve rare books by scanning them without removing their spines.