A settlement agreement formally approved by a U.S. federal court in July 2026 has set the highest financial benchmark yet in the AI industry’s copyright disputes: Anthropic will pay $1.5 billion to the rights holders of about 500,000 books, making it the largest copyright settlement in U.S. history. But the money has not quelled everyone’s anger.
7 Million Pirated Copies, 7 Million Points of Infringement
In August 2024, three authors—Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson—filed a copyright lawsuit against Anthropic, alleging that the company had downloaded books in bulk from “shadow libraries” to train its Claude models. Court filings later disclosed that Anthropic had downloaded more than 7 million book copies from two piracy databases, Library Genesis (LibGen) and Pirate Library Mirror (PiLiMi).
At Stanford University, the impact of the lawsuit was especially concentrated. According to an October 27, 2026 report by The Stanford Daily, the works of at least six Stanford professors were confirmed to have been pirated for model training, including Tobias Wolff, professor emeritus in the English department; medical humanities writer Abraham Verghese; novelists Chang-Rae Lee and Daniel Mason; and business school professors Hayagreeva Rao and Myra Strober. Wolff alone had 10 works included in the training set without authorization.
The Judge’s “Dividing Line”: One Act, Two Legal Outcomes
In June 2025, Judge William Alsup of the federal court for the Northern District of California issued a subtly worded but far-reaching ruling: Anthropic’s use of legally purchased books for AI training was “classic transformative use” protected by the fair use doctrine, but downloading and storing copies of those books from piracy sites such as LibGen was not protected by fair use and constituted direct infringement.
Alsup wrote in his ruling: “Like any reader aspiring to write, Anthropic’s large language models trained on these works not to surpass and copy them—but to turn a corner and create something different.” The analogy equated AI reading books with human readers internalizing style and themes, providing support for the fair use argument.
However, the legal line between “reading books” and “stealing books” was thus drawn quite clearly: AI training itself may be protected, but the means of obtaining the raw material determines where legal liability lies. Anthropic stepped right on that line—it not only bought books, but also stole them.
$1.5 Billion Settlement Covers 500,000 Works, About $3,000 Each on Average
In August 2025, Anthropic reached a $1.5 billion class settlement with the plaintiffs. It received preliminary approval on September 25, 2025, and final approval was issued by Judge Martínez-Olguín on July 20, 2026. According to an analysis by Authors Alliance, the settlement covers about 500,000 eligible copyrighted works, with average compensation of about $3,000 per work. Attorneys’ fees were originally requested at $187.5 million but were ultimately reduced by the court to about $101.6 million.
That figure is the largest single settlement in U.S. copyright history. But it also exposes an order-of-magnitude gap: from 7 million pirated downloads to 500,000 valid claims, and then to $3,000 per work—every step in the calculation chain dilutes the compensation that an individual creator can recover.
28 Authors Reject the Settlement and File a Separate Suit Seeking Up to $150,000 per Work
Tobias Wolff made no secret of his dissatisfaction with the settlement. “Forcing me to work for them in this way without paying me a cent is a form of slave labor,” he told The Stanford Daily. “They are using my work without paying anything. Over time, if writing does not pay, they will make writers’ lives extremely difficult.”
In May 2026, Wolff, Chang-Rae Lee, and 28 authors who opted out of the class settlement jointly filed a separate lawsuit, Cruz v. Anthropic, requesting a jury trial and seeking statutory damages of up to $150,000 per work—about 50 times the per-work compensation in the class settlement. They also asked the court to issue a permanent injunction barring Anthropic from continuing to use their works in future model training.
According to PACER court docket tracking, Cruz v. Anthropic was filed on May 13, 2026, in the U.S. District Court for the Northern District of California. The case remains at an early stage, with no substantive ruling yet.
This Is Not Just Anthropic’s Problem
Broadening the view, Anthropic is not the only AI company under heavy pressure over training data compliance. Meta has been sued by multiple authors over its use of the Books3 dataset. A federal judge ruled in 2026 that its training conduct constituted fair use, but Meta still faces ongoing infringement claims because it “seeded” pirated files to other users during the downloading process. OpenAI has been locked in copyright class action litigation with The New York Times and several news organizations; in September 2026, the two sides filed motions for summary judgment, and the outcome has not yet been announced.
Two California federal judges both ruled in mid-2025 that using copyrighted books to train large language models can in principle constitute fair use. That direction has provided some legal cover for the AI industry’s training practices in the short term. But another principle established in the Bartz case is equally impossible to ignore: the means of acquisition is itself an independent legal issue. Fair use protects the act of training, not the act of downloading pirated material.
The “Transaction Cost” Dilemma of Licensing
Why would AI companies go through piracy repositories instead of buying licenses directly from copyright holders? Stanford Law School professor Paul Goldstein—himself one of the plaintiffs in the Bartz class action—pointed to the “transaction cost” problem in court-related materials: negotiating licenses for millions of works one by one is technically extremely difficult, and the infrastructure for a licensing system simply does not exist.
Mary Rasenberger, Executive Director of the Authors Guild, offered what she considers a workable solution: a licensing mechanism would not only compensate authors but could also require filtering mechanisms for AI outputs. The guild is currently working with Congress to promote national licensing legislation, seeking to fill this gap at the institutional level.
That path is reasonable in theory, but the obstacles at the implementation level are obvious: Who would operate a unified licensing platform? Who would bear the burden of confirming copyright ownership? How would cross-border copyright be handled? The final distribution of damages in the Bartz case has taken more than a year and covers only 500,000 of the 7 million copies—licensing millions of works would be another order of magnitude in complexity.
A Business Choice Confirmed by Court Ruling
In hindsight, Anthropic’s conduct reveals a rational calculation: in the absence of mature licensing infrastructure, the legal risk of obtaining training data in bulk from piracy databases—even in the worst case, such as the $1.5 billion in the Bartz case—may still be lower than the combined cost of obtaining millions of copyrighted works through compliant means. This was not a technical error, but a business decision priced after the fact through court proceedings.
$1.5 billion is the largest copyright payment to date, but compared with the scale of financing and market valuation Anthropic has already achieved in the large-model race, it remains a commercially digestible cost. What the 28 authors in the Cruz case are pursuing is a statutory damages standard of up to $150,000 per work, to force this calculation to be redone.
If the Cruz case ultimately succeeds, its chain effects will extend far beyond these 28 authors: it would establish a benchmark—rights holders who opt out of a class settlement can obtain compensation close to the statutory cap through individual litigation. That would fundamentally change the bargaining logic of AI companies in data compliance negotiations.
For the entire industry, the real legacy of the Bartz case is not the $1.5 billion, but the line Judge Alsup drew: training can be fair use, but downloading pirated material cannot. The next question is whether the parties can build a licensing system that actually works on the lawful side of that line—rather than waiting for the next, more expensive ruling to force an answer to the question.
Sources: - [Bartz v. Anthropic Settlement: Understanding America's Largest Copyright Settlement](https://legalblogs.wolterskluwer.com/copyright-blog/the-bartz-v-anthropic-settlement-understanding-americas-largest-copyright-settlement/) - [Authors Secure $1.5 Billion Settlement in Landmark AI Piracy Case](https://www.lieffcabraser.com/2025/09/authors-secure-1-5-billion-settlement-in-landmark-ai-piracy-case/) - [Anthropic illegally trained Claude with Stanford professors' books – Stanford Daily](https://stanforddaily.com/2026/10/27/professors-books-pirated-to-train-claude/) - [Cruz et al v. Anthropic PBC – PACER Monitor](https://www.pacermonitor.com/public/case/64645512/Cruz_et_al_v_Anthropic_PBC) - [AI Training Data Copyright Lawsuits (2026)](https://ailawsuittracker.com/issues/training-data-copyright/)© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接