Court Documents Reveal Anthropic's 'Project Panama' to Destructively Scan Books for Claude Training
Key Takeaways
- ▸Anthropic's internal 'Project Panama' systematically destructively scanned physical books to train Claude, bypassing copyright permissions
- ▸The company prioritized logistical solutions over legal compliance, initially using pirated sources before switching to destructive scanning after legal concerns arose
- ▸A court ruled that using copyrighted works to train AI models constitutes 'fair use' and 'transformative use,' setting precedent for how AI companies can legally source training data
Summary
A California court case decided in late July 2026 revealed that Anthropic undertook "Project Panama," an effort to destructively scan printed books to obtain training data for its Claude language model, rather than securing copyright permissions from authors. The practice—which involved physically scanning books and then destroying the originals—was initially prompted by Anthropic's desire to avoid the "legal/practice/business slog" of managing copyright with book owners. After initially attempting to use pirated sources, the company pivoted to destructive scanning as a "transformative" use of copyrighted materials.
The Bartz v Anthropic PBC case resulted in a $1.5 billion out-of-court settlement with authors. However, the presiding judge ruled that using proprietary material to train a language model does not itself constitute copyright infringement, treating AI training as analogous to human learning. The decision hinged on whether the transformation of printed books into digital datasets and their subsequent destruction qualified as "fair use" under US law—a determination with significant implications for how AI companies source training data.
The revelation raises critical questions about the sustainability and ethics of scaling generative AI, particularly whether corporations should be permitted to consume vast amounts of human-created cultural works without compensation or consent from their authors.
- The $1.5 billion settlement with authors and the court's decision underscore the tension between AI development and intellectual property rights
Editorial Opinion
The revelation of Project Panama exposes a troubling pattern: Anthropic viewed legal negotiation with authors as a bothersome obstacle to overcome, rather than a necessary partnership. While the court's ruling treats AI training as equivalent to human learning, the industrial scale and commercial intent of companies like Anthropic fundamentally differ from an individual reader. Permitting corporations to extract and destroy entire libraries of human creativity without consent or compensation risks reducing authors to mere fuel for AI engines rather than partners in knowledge creation.
