Why Anthropic Bought and Destroyed Millions of Books to Train AI

Summary

Court records and investigation reports reveal that Anthropic purchased millions of second-hand books for its 'Project Panama' initiative, where the bindings were cut, every page was scanned, and the paper was recycled to train Claude models. This practice triggered a landmark case in copyright law; separately, a multi-billion dollar settlement was reached regarding copies acquired through pirated channels.

Why it matters

This case establishes one of the most significant precedents in AI copyright law. A U.S. federal court ruled that scanning legally purchased books using destructive methods could be considered fair use. However, it explicitly stated that retaining pirated digital copies does not fall under this protection. As the internet becomes saturated with AI-generated text, the strategic pursuit of 'clean' human-written books is reshaping the publishing economy and the discourse on preserving cultural heritage. This situation has implications for all major AI laboratories.

Highlights

• Anthropic hired Tom Turvey, a former Google Books executive. The company used industrial hydraulic cutters to remove bindings before high-speed scanning, after which the paper was recycled.
• Judge William Alsup ruled that scanning legally purchased books for AI training could qualify as fair use. However, a separate ruling determined that the company's pirated book archive was outside this protection. This process resulted in a multi-billion dollar settlement.
• According to the article, Meta, Google, OpenAI, and unnamed intermediaries engaged in similar bulk book purchases. These companies sometimes preferred out-of-print works to avoid drawing attention to themselves.

Critical take

The article was published in an architecture magazine. The text places excessive emphasis on the 'destroyed books' framing while downplaying the court's fair use ruling for legally purchased copies. Additionally, it blends verified court records with unverified claims about other companies (using phrases like 'allegedly' or 'according to reports') without clear distinction. This approach shifts the piece closer to opinion than investigative journalism.