Judge approves Anthropic's 1.5 billion dollar book piracy settlement

Judge Araceli Martinez-Olguin approved Anthropic's 1.5 billion dollar settlement with authors on July 20, the largest copyright payout in US history, according to TechCrunch. She overruled objections that argued the sum was too small.

The payout size is the headline, but the mechanism is the part that should worry other AI companies. Anthropic has to destroy more than 7 million pirated book files that Judge Alsup found it had unlawfully stored, and the per-work payout runs four times the statutory minimum.

The deal pays roughly 3,000 dollars per work across more than 480,000 books, against a 750 dollar statutory minimum. Some authors opted out of the class and are litigating separately, according to TechCrunch.

What the settlement doesn't do is settle the question everyone in AI keeps asking: is training on copyrighted text legal? Alsup already answered that.

He separately ruled that Anthropic's training on the books was fair use. That ruling sets no binding precedent from this case, since the settlement resolved the piracy and storage claims, not the training question.

The distinction is the whole story. Anthropic paid 1.5 billion dollars because it pirated and stored 7 million book files without a license. It didn't pay for training a model on copyrighted text, because a judge already said that part was legal.

If you're building on scraped or aggregated data, that's the line worth watching. Acquisition and retention carry the legal risk here. How you got the data and whether you kept it matters more right now than how a model learned from it.