News analysis
Anthropic's $1.5 Billion Settlement Draws a Line Between AI Training and Piracy
Anthropic's $1.5 billion copyright settlement covers pirated books — not a ruling that AI training itself is illegal. Here is the line the case actually draws.
By Ongoing AI · Published July 27, 2026 · Last updated July 27, 2026
A federal judge has given final approval to Anthropic's $1.5 billion settlement with authors and publishers whose books appeared in a collection assembled from piracy sites. The amount is historic. The legal distinction behind it is more important.
The court did not rule that training an AI model on copyrighted books is inherently illegal. In an earlier decision, Judge William Alsup found that Anthropic's use of books to train its Claude models was sufficiently transformative to qualify as fair use. The unresolved liability came from how Anthropic acquired and retained the books.
Anthropic downloaded more than seven million books from sources including Library Genesis and Pirate Library Mirror, according to the litigation, and stored them in a permanent central library. The court treated that library as a separate use from model training, and found that piracy could not be excused merely because some books might later support a transformative process. The resulting settlement establishes a consequential principle: what a company does with training data and how it obtained that data are separate legal questions.
What happened
On July 20, US District Judge Araceli Martínez-Olguín granted final approval to the settlement in Bartz v. Anthropic and entered judgment. Anthropic will provide a non-reversionary fund of at least $1.5 billion covering 482,460 works — roughly $3,000 per work before legal fees, expenses, and allocation among competing rightsholders. As of April 16, claims had been submitted for 440,490 works (91.3% of the final list), with 350 valid opt-outs covering 1,802 works and 54 objections or comments.
The settlement also requires Anthropic to destroy original files downloaded from Library Genesis and Pirate Library Mirror, along with copies originating from them, subject to legal-preservation requirements. Anthropic has denied wrongdoing; the settlement resolves the class's covered claims without a trial or admission of liability. Judge Martínez-Olguín approved approximately $101.6 million in attorneys' fees — well below the $187.5 million the plaintiffs' lawyers most recently requested and their earlier $300 million proposal.
Timeline
Authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson sue Anthropic, accusing it of copying copyrighted books without permission while developing Claude.
Judge William Alsup rules that using books to train Claude was transformative fair use, but that maintaining a permanent general-purpose library of pirated books was not justified.
Anthropic and the certified class sign a settlement term sheet resolving the remaining piracy claims before a scheduled December trial where statutory damages could have run far higher.
Judge Alsup grants preliminary approval, finding the agreement sufficiently fair to proceed through notice, claims, and final review.
The settlement administrator reports valid claims covering 440,490 of the 482,460 works on the final list — 91.3% — with 350 opt-outs and 54 objections or comments.
Judge Araceli Martínez-Olguín holds the final approval hearing, considering the settlement, objections, opt-out disputes, and the attorneys' fee request.
Judge Martínez-Olguín grants final approval and enters judgment, describing the $1.5 billion settlement as meaningful relief in a novel and risky case.
What is confirmed — and what is not
The settlement fund ($1.5B minimum, 482,460 works), the ~91.3% claims rate, the ~$3,000-per-work gross estimate, the destruction requirement, the ~$101.6M fee award, and the absence of any admission of liability are all confirmed by the court's order. The court previously decided that training on the books was transformative fair use, that digitizing lawfully purchased print books was fair use, and that a permanent library of pirated books was not.
What is not established is broad: that every use of copyrighted material for AI training is fair use; that this ruling binds other courts; that any particular Claude output infringed; that ~$3,000 sets a licensing rate; or that other AI companies using similar datasets will reach the same outcome. The fair-use decision came from one district court on a specific record. Other judges remain free to analyze different models, datasets, and market effects differently.
Why operators should care
The following is Ongoing AI analysis.
The central lesson is that recording what entered a training corpus is not enough — you need to know where each item came from and what rights accompanied it. Two identical digital books create different legal risk if one was lawfully purchased and the other downloaded from a piracy site. A credible AI data-provenance system should be able to answer, for any file: where and when it was acquired; whether it was purchased, licensed, public, or scraped; which agreement covered its use; whether the source was authorized to distribute it; which models, evaluations, or libraries received it; and whether every derived copy can be identified and deleted.
The case also shows that a general data library can create its own liability. Anthropic argued its book collection supported model development, but the court separated the permanent library from the later training — a transformative downstream use may not cure an unlawful upstream acquisition. Developers should distinguish raw acquired materials, general research libraries, curated training corpora, evaluation datasets, and derived artifacts, rather than treating all five as one undifferentiated "training data" system. And deletion must propagate: fulfilling a destruction order means locating extracted copies, processed versions, deduplicated records, shards, backups, and assembled datasets — without reliable lineage, proving deletion is difficult.
For enterprise buyers, none of this implies that Claude users are liable for Anthropic's training practices. It does add a category of vendor diligence: how a provider documents acquisition, whether it separates licensed, public, and disputed sources, what indemnification covers outputs, and whether material litigation could disrupt models — questions that matter most in publishing, media, education, and other copyright-intensive industries.
Why investors should care
The following is Ongoing AI analysis.
The $1.5 billion is frequently described as a penalty; legally it is a negotiated class-action settlement. Anthropic admitted no wrongdoing and the case did not reach a damages verdict — the payment removed the risk of a trial in which statutory damages could potentially have reached hundreds of billions. That distinction matters when comparing this with government enforcement or a final finding of willful infringement.
The cash obligation is staged across four payments — $300M (October 2025), $300M (shortly after approval), $450M (September 2026), and $450M (September 2027) — which reduces the immediate burden but leaves substantial scheduled payments into next year. Strategically, the outcome cuts both ways: the fair-use ruling gave AI companies a strong argument that training can be transformative without individual licenses, while the piracy ruling warned that a transformative use does not excuse unlawful acquisition. Copyright risk therefore can't be estimated solely by asking whether training is fair use; a company's data-collection history and retention practices may be just as consequential.
Finally, ~$3,000 per work is not an industry licensing price. It reflects litigation risk, statutory-damage exposure, class certification, the alleged method of acquisition, and the value of ending the case before trial — a negotiated license for future training could be structured entirely differently. And the agreement is deliberately limited: it does not release conduct after August 25, 2025, works outside the final list, output-infringement claims, opt-out claims, or claims against other companies. The industry's wider copyright exposure remains unresolved.
Companies, people, and legal structures involved
Claude — Anthropic's family of AI models and products; the training uses examined in the case were found to be transformative fair use.
Bartz v. Anthropic — the class action in the US District Court for the Northern District of California, case number 3:24-cv-05417. Named plaintiffs include novelist Andrea Bartz (We Were Never Here), nonfiction author Charles Graeber (The Good Nurse), and Kirk Wallace Johnson (The Feather Thief).
Judge William Alsup — issued the June 2025 fair-use ruling and granted preliminary approval before retiring. Judge Araceli Martínez-Olguín — conducted the final review, approved the settlement, and entered judgment on July 20.
Library Genesis and Pirate Library Mirror — the shadow-library sources covered by the settlement's destruction requirement.
What to watch next
Final approval closes one major stage, but several developments remain.
- Whether any appeal delays the settlement's effective date, and when distributions to authors begin.
- The final per-work payment after fees, expenses, and allocation.
- Certification that covered piracy-derived files have been destroyed.
- Anthropic's remaining ~$900 million in scheduled payments (September 2026 and 2027).
- Separate lawsuits by authors and publishers who opted out, and new claims over conduct after August 25, 2025.
- Litigation alleging infringing model outputs, and how other courts apply fair use to different training records.
- Whether insurers, investors, and enterprise customers begin requiring auditable dataset lineage.
The settlement does not answer the broadest question in AI copyright law: it does not decide that all training is fair use, nor that all unlicensed training is infringement. It draws a narrower, immediately practical line — transformative model training may receive fair-use protection, but building that model from a poorly documented library of pirated material can still generate enormous liability. For AI companies, the provenance of the data may now matter as much as the purpose of the model.
Sources and updates
- Final approval order in Bartz v. Anthropic — US District Court, N.D. Cal., July 20, 2026 · primary source · accessed July 27, 2026
- Official settlement documents — Anthropic Copyright Settlement administrator · primary source · accessed July 27, 2026
- Official settlement FAQ — Anthropic Copyright Settlement administrator · primary source · accessed July 27, 2026
- Order on fair use (June 2025) — US District Court, N.D. Cal., June 23, 2025 · primary source · accessed July 27, 2026
- US judge approves Anthropic's $1.5 billion settlement in copyright lawsuit — Reuters, July 20, 2026 · accessed July 27, 2026
- Landmark Anthropic settlement: judge rejects windfall for lawyers — Reuters, July 21, 2026 · accessed July 27, 2026
- Court grants final approval of Anthropic copyright settlement — Authors Guild · accessed July 27, 2026
- Associated Press report on the Anthropic settlement — Associated Press · accessed July 27, 2026
Update history
- July 27, 2026 — Initial publication.
Get the next briefing
Ongoing AI tracks the companies, capital, technology, and enterprise adoption shaping the AI economy.
The consequential AI developments — connected to the companies, capital, technology, and enterprise adoption they affect. Unsubscribe anytime.