AI Copyright Case No. 1 Settles: Anthropic Pays Authors $1.5 Billion
In September 2025 Anthropic agreed to pay about $3,000 per book across ~500,000 books — the first major settlement over AI training data.
Per AP and NPR, Anthropic agreed on September 5, 2025 to pay $1.5 billion to settle a class action by authors who accused it of training Claude on pirated books. The payout covers roughly 500,000 books at about $3,000 each; a federal judge granted preliminary approval on September 25.
It was the first time an AI company paid massive damages over training data — dubbed 'AI copyright case No. 1.'
The Key Ruling: Where Fair Use Ends
The court's earlier distinction is the case's real precedent value: training on legally purchased books is fair use, but downloading pirate libraries is not. That refined the fight from 'can you train on copyrighted content' to 'how did you get it' — lawful acquisition makes training defensible; illegal sourcing has no shelter. The industry's compliance line has been bright ever since.
The Payout Structure: What $3,000 a Book Signals
Around 500,000 books at roughly $3,000 each — a unit price far above retail is effectively punitive pricing: it sets the cost of retroactive make-goods well above upfront licensing. For anyone training new models, the rational choice flipped from gambling on litigation to buying licenses — which is why the settlement matters more than a verdict: it repriced the risk.
The Twin Storyline With NYT v. OpenAI
The year's other thread was NYT v. OpenAI clearing dismissal (see our coverage): Anthropic chose to pay and close the door; OpenAI chose to fight in court — the contrast between the two strategies will shape training-data rules for the next decade. Content owners' leverage rose after both cases, tipping the licensing market from buyers to sellers.
Our Take
A federal judge granted preliminary approval on September 25. The settlement rewrote training-data economics: licensed content went from optional to mandatory, and data cost formally entered the LLM cost sheet. Long term that may be healthy: when data must be paid for, high-quality content production finally earns cash flow from the AI era — the turning point where the content and AI industries move from confrontation to commerce.
This article aggregates official announcements and public reporting; original sources are linked below.
Source:美联社