What the Anthropic $1.5B settlement really means for AI companies

The August 2025 announcement that Anthropic would pay $1.5 billion to settle Bartz v. Anthropic was the largest copyright settlement in United States history. It was also the first time a major AI lab paid out at scale on training-data claims. Reporting at the time framed the deal as a watershed; six months later, with the opt-out window closed and the final fairness hearing approaching, the more useful question is narrower: what does Bartz actually settle, and what does it leave open?

The doctrinal hinge

Judge William Alsup's June 2025 order in Bartz did something subtle. It did not hold that training large language models on books is unlawful. It held that downloading and retaining pirated copies of books from shadow libraries — Library Genesis, Pirate Library Mirror, and the Books3 corpus — was not protected by fair use, even when the downloaded copies were used (or intended to be used) for training. Training on legitimately acquired books, the court said, may well be transformative.

The opinion's most-quoted passage is also the operative one for litigation strategy. Anthropic, the court found, "could have lawfully purchased the books" and chose instead to "steal them" to avoid what its CEO had called the "legal/practice/business slog" of licensing. That language matters because it locates the unlawful conduct upstream of the training process itself. Anthropic's exposure was not for what its model learned. It was for how its corpus was acquired.

Why $1.5 billion, and why so fast

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 low end of statutory damages — say $750 per work, the floor — the math reaches eight or nine figures quickly. At the high end, a class-aggregate verdict would have been an existential threat. The settlement worked out to roughly $3,000 per work for a class of about 500,000 authors. That is a number that bears almost no resemblance to a fair-use ruling and substantial resemblance to a settlement struck in the shadow of statutory damages.

This explains why the deal closed so quickly after Alsup's order. Anthropic could not reasonably price the upside of going to trial against the cost of a verdict that, if it came in for plaintiffs, would have produced existential exposure. Settlement was rational risk management; it was not — and the court did not treat it as — a concession that the underlying training was infringing.

The floor, not the ceiling

Three implications now follow that every AI defendant's general counsel should treat as decided.

First, the per-work damages anchor is set. $3,000 per work, multiplied by class size, is the number every plaintiffs' firm in every pending books case will quote. Kadrey v. Meta, the Concord Music sister action, the OpenAI books cases coming after Authors Guild — all of them now have a ready answer when judges ask about damages models. The floor is $3,000. The ceiling, if a case goes to verdict on willful infringement, is much higher.

Second, the plaintiffs' bar will multiply. Class counsel in Bartz has petitioned for $300 million in fees — twenty percent of the fund, supported by roughly 26,000 hours of attorney work. Those fee economics are unusually attractive. The pipeline of follow-on filings will not slow.

Third — and most important strategically — the legality of training itself remains unresolved at the appellate level. Bartz settled on the way to trial. Alsup's bifurcated order is a district-court opinion, not Ninth Circuit law. The cleanest fair-use ruling on AI training to date in the United States remains Thomson Reuters v. Ross Intelligence (D. Del. 2025), which went the other way — finding that training on copyrighted material to build a competing product was not fair use. The doctrinal conflict is real.

What labs should actually do

The operational implications run in three directions. Audit the corpus: every lab is now obligated, as a matter of legal hygiene, to be able to certify whether a given training set includes pirated material. Plaintiffs' counsel will ask in every case. Document provenance: licensing receipts, terms-of-service captures, and crawl logs are the records that will determine whether a lab's defense looks like Anthropic's pre-Alsup posture or something more defensible. And model what Bartz means for fine-tuning: the order's logic — that unlawful acquisition is not laundered by transformative use — applies as forcefully to fine-tuning corpora as it does to base-model pretraining.

The headline, stripped of drama, is small and precise. Bartz did not settle whether AI training is fair use. It settled what happens when an AI lab cannot prove it acquired its training corpus lawfully. That is a narrower holding than the $1.5B sticker suggested. It is also a hardier one, and the next year of litigation will be conducted under it.