Anthropic just paid $1.5 billion
for the claim it lost.
No court will ever be asked
to check the claim it won.
On July 20, a federal court finalized the largest copyright settlement in history — $1.5 billion for pirating books. The ruling that actually matters closed thirteen months earlier, was never appealed, and just became permanently unreviewable — which is already being blurred with “settled law,” including by the company that benefits most from the confusion.
A federal court signed the largest copyright settlement in history on July 20 — $1.5 billion, roughly 500,000 books, about $3,000 a title, in Bartz v. Anthropic PBC. Almost every outlet led with that number — real, and the wrong headline. It is the fine for the part of the case everyone already knew Anthropic would lose: downloading millions of books from pirate libraries. The ruling that will actually matter to the next company sued over AI training came thirteen months earlier, in the same docket, from a judge who never saw how the case ended. Training a language model on copyrighted books, he held, is fair use. That finding was never appealed. As of July 20, it never can be.
Judge William Alsup made that call in June 2025, on summary judgment. Training Claude on the books, he wrote, was “quintessentially transformative”: Anthropic’s models trained on the works “not to race ahead and replicate or supplant them, but to turn a hard corner and create something different.” Digitizing the purchased print copies was fair use too — a format change, not a new copy. On both questions, Anthropic won outright. On the third — the books pulled from pirate libraries like LibGen to build a permanent training archive — Alsup drew the opposite line: “That Anthropic later bought a copy of a book it earlier stole off the internet will not absolve it of liability for the theft, but it may affect the extent of statutory damages.” Only that claim turned into $1.5 billion instead of a verdict.
Alsup never signed off on the settlement. He retired at the end of 2025; the case passed to Judge Araceli Martínez-Olguín, who signed the final approval on July 20 — two judges, thirteen months apart, closing two different questions. Had the piracy claim gone to trial or appeal, the fair-use ruling would have traveled with it, exposed to reversal. Settling instead ended the case outright: dismissed with prejudice, no appeal left standing. Anthropic’s deputy general counsel, Aparna Sridhar, called the fair-use finding a ruling that “remains the law today.” It isn’t, strictly — a single district court’s opinion binds no other court. What it now is, permanently, is unchecked. That gap is doing real work in every headline about this case this week.
What actually closed
The fair-use question is closed to review, not settled as precedent. Training an LLM on books, and digitizing lawfully purchased copies to build the training set, both stand as fair use in this docket — unappealed, now beyond any appeal, and not binding law for the next company sued on the same theory.
The piracy question was closed by a check, not a ruling. The $1.5 billion, the $3,000 per work, the 91.3 percent claimed as of April 16 — all of it resolves damages for this case. None of it tells the next AI company what happens if it gets caught building a pirated library instead of licensing one.
A different judge already flagged exactly this gap. Four months before Alsup ruled, in Thomson Reuters v. Ross Intelligence, Judge Stephanos Bibas rejected fair use for a non-generative search tool trained on Westlaw’s headnotes — and wrote his own caveat: “the AI landscape is changing rapidly, I note for readers that only non-generative AI is before me today.” Alsup’s ruling is the only answer on record to that question — and, thanks to the settlement, the only one that will ever be tested.
What it settles nothing about
Two rulings, one case, one year: the line between a model allowed to learn from a work and a model allowed to compete with the market for it. What it doesn’t touch is what happens after training is legal — whether the model, once trained, knows who you are. Fair use decides whether Claude was allowed to read the material. It says nothing about whether Claude, or any answer engine, can find you inside it.
A model can be perfectly and permanently trained on every credential you have ever published and still misclassify you, fragment your signals, or default to whichever competitor happens to be structured so the model can read it. That failure had a name before this settlement existed: Digital Derangement Syndrome™ — a recognition failure, not a training one. No court has been asked to rule on it, because it was never a legal question. Identity Architecture™ answers the question this case never asked: not whether a model was allowed to learn, but whether it was built to recognize you once it had.
Nineteen editions in, the pattern holds again. The institutions catching up to this era keep settling what a model is allowed to do, and keep leaving untouched what it actually sees. Anthropic just spent $1.5 billion and thirteen months making sure one favorable opinion can never be checked. Nothing in that record gets your name recognized inside what it was just permanently allowed to build.
The Court’s Order Granting Final Approval of Class Action Settlement, Bartz v. Anthropic PBC, was filed July 20, 2026 (Case No. 3:24-cv-05417-AMO, N.D. Cal.) — case docket, CourtListener.
Judge Alsup’s June 23, 2025 order on fair use, Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025), is broken down with the order’s own language by Loeb & Loeb and Good Journey Consulting.
Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., 2025 WL 458520 (D. Del. Feb. 11, 2025), the non-generative-AI ruling Alsup’s decision effectively answers, is analyzed by Davis Wright Tremaine.
Anthropic’s statement on the final approval, from deputy general counsel Aparna Sridhar, is reported by Benzinga; the non-binding-precedent read is reported by Tech Times.
Judge Alsup’s December 31, 2025 retirement and the case’s reassignment to Judge Araceli Martínez-Olguín are reported here.
The Authors Guild’s response to final approval is on the record at authorsguild.org.