Anthropic pays $1.5B for pirated training data

A judge approved a $1.5B settlement between Anthropic and authors whose books were used without license to train Claude. The HN comments were mostly dark humor, 'that number is missing a zero or two in front of the decimal point,' plus jokes about paying in expiring Fable credits. The legal framing matters: the settlement is specifically about piracy, not about whether training on books is permissible in principle.

This is the first big closed settlement of the wave of AI copyright litigation, and it sets a number. Whether that number is low or high relative to actual harm is debatable, but it exists now. Every lab training on web-scraped or ambiguously licensed data has a data point to anchor their legal exposure estimates.

The broader pattern: training data provenance is becoming a legal liability, not just an ethical footnote. The labs that built clean data pipelines early are in a structurally better position than those who scraped first and asked questions later.