Bartz v. Anthropic settlement explained
In 2025, Anthropic agreed to pay $1.5 billion to settle a books-class action in the Northern District of California. It became the largest copyright settlement in US history and the first major AI-training payout. Here is what it actually decided.
Bartz v. Anthropic was filed in 2024 by author Andrea Bartz and a class of writers whose books appeared in the Books3 corpus and other shadow-library datasets used to train Claude. The case proceeded before Judge William Alsup in the Northern District of California — the same judge who handled Oracle v. Google.
What Judge Alsup actually held
In June 2025, Judge Alsup issued a bifurcated summary-judgment ruling. The opinion drew a sharp line. Training a large language model on books, Alsup wrote, may well be transformative. But the act of downloading and retaining pirated copies from Library Genesis, Pirate Library Mirror, and similar sources — even if the downloaded copies were used for training — was not protected by fair use.
The most-quoted passage is the operative one: Anthropic, the court found, "could have lawfully purchased the books" and chose instead to "steal them." That phrase locates the unlawful conduct upstream of training itself. The exposure was not for what Claude learned; it was for how the corpus was acquired.
Why $1.5 billion
Once that order issued, Anthropic faced a willful-infringement jury trial with statutory damages of up to $150,000 per work across hundreds of thousands of pirated titles. Even at the floor — $750 per work — the math reaches eight or nine figures. At the ceiling, an existential verdict was on the table.
The settlement worked out to roughly $3,000 per work for a class of about 500,000 authors. Class counsel petitioned for $300 million in fees on top of the settlement fund — about 20%, supported by approximately 26,000 attorney hours.
What the settlement decided — and didn't
The settlement resolved Anthropic's exposure on the underlying claim. It did not produce an appellate fair-use ruling. The cleanest US fair-use opinion on AI training to date remains Thomson Reuters v. Ross Intelligence in D. Del. (2025), which went the other way. The doctrinal conflict is real and unresolved.
What the settlement did do is set a damages anchor. $3,000 per work, multiplied by class size, is now the number every plaintiffs' firm in every pending books case will quote. That number will appear in Authors Guild v. OpenAI and in every books class action that survives summary judgment. Notably, the Bartz number did not save the Kadrey plaintiffs — Meta won training-stage fair-use summary judgment in June 2025 on a thin market-harm record.
What every AI lab should do now
Operationally: audit the corpus, document provenance, and apply the same logic to fine-tuning datasets that Bartz applied to base-model pretraining. The order's holding — that unlawful acquisition is not laundered by transformative use — applies as forcefully to fine-tuning as it does to pretraining.
Frequently asked questions
How much did Anthropic pay in the Bartz settlement?
$1.5 billion — the largest copyright settlement in US history, and the largest publicly disclosed AI training settlement to date.
Who gets paid from the Bartz settlement?
Authors in the certified class — roughly 500,000 writers whose books appeared in the pirated training corpora. The deal works out to about $3,000 per work.
Does the Bartz settlement mean AI training is illegal?
No. The settlement resolved Anthropic's exposure for using pirated training data. The judge's underlying order suggested training itself may be transformative — the unlawful conduct was the use of pirated source material, not the act of training.
Will the Bartz settlement affect other AI lawsuits?
Yes. The $3,000-per-work damages anchor will be cited in every subsequent books class — the Authors Guild MDL and the Concord Music sister actions. The number will not, however, rescue cases that lose at summary judgment: Meta won training-stage fair-use SJ in Kadrey v. Meta two days after the Bartz ruling.