Bartz v. Anthropic and Copyright Infringement in the Age of AI
By Lucas Yanney ’29
Since their emergence, generative Artificial Intelligence (AI) applications like OpenAI’s ChatGPT and Google’s Gemini have been highly controversial due to their potential to displace millions of jobs and the degradation of human-to-human interaction. Additionally, AI models “hallucinate,” or fabricate false information—a phenomenon that has been prevalent in the legal world, as throughout the past year, law firms have cited fictitious cases in their court filings. (1) However, the risks of reliance on AI chatbots extend even deeper into the legal system, beyond attorneys’ misuse of chatbots. Claims of copyright infringement lie at the heart of the prevailing class-action lawsuit filed against Anthropic, the developer of the Large Language Model (LLM) Claude. The plaintiffs—authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson—represent a certified class of authors whose works were among approximately 7 million copies of books that are alleged to have been downloaded from unauthorized sources to be used in the training of their AI models. (2)
Like many generative AI companies, Anthropic relies on extensive datasets to train its models to recognize patterns and generate human-like responses. The plaintiffs in Bartz v. Anthropic argue that this process necessarily involves the reproduction of protected works which allow the model to learn from considerable portions of copyrighted material. The case revolves around Anthropic’s unauthorized use of several million copies of works from the piracy website Library Genesis, or LibGen, along with other online “shadow libraries.” (2) The underlying question raised in this lawsuit is whether the training of an AI model using copyrighted works constitutes permissible use under existing copyright law or, instead, amounts to large-scale infringement. (3)
While Judge Alsup’s fair use analysis highlights the doctrinal divide regarding LLM training, the procedural history itself of Bartz is equally significant. Unlike most copyright infringement cases, where the lawsuit is brought by an individual rights holder, Bartz proceeded as a class action lawsuit, whereby 367,824 unique authors are represented in a single, certified class of copyright holders. (4) Rule 23 of the Federal Rules of Civil Procedure outlines the requirements for class action lawsuits and the certification of the plaintiff class in such a suit. A class action lawsuit is a legal or procedural mechanism where all potential claims against one defendant can be combined into a single lawsuit. For a class of plaintiffs to be certified, there are four prerequisites: numerosity, commonality, typicality, and adequacy. (5) In the majority of copyright infringement cases, however, these requirements are not met, since claims often depend on the individualized assessments of liability and fair use. Nevertheless, the nature of generative AI model training alters this traditional framework. Because LLMs are uniformly trained on immense datasets, the alleged harm is largely the same to each of the copyright holders. As a result, Bartz represents a rare instance of a class action copyright infringement lawsuit.
The central framework of this argument is the doctrine of fair use, outlined in 17 U.S.C. § 107, declaring the use of copyrighted work without authorization permissible under certain circumstances. (6) The four factors that courts use to evaluate the fair use of copyrighted materials are: purpose and character, nature, amount and substantiality, and the effect that the use has on the potential market. (7) However, the question of fair use becomes all the more complicated in the context of copyrighted materials used in generative AI applications. It can be argued that the use is transformative, and thus not copyright infringement, as models do not publish words verbatim but rather extract relationships to generate new outputs, through analysis of the statistical patterns between strings of words in a sentence or paragraph; however, the plaintiffs assert that the scale of copying (often entire works) blatantly defies fair use, especially when the training data is sourced from pirated repositories. (8)
On June 23, 2025, Judge William Alsup of the U.S. District Court for the Northern District of California issued a preliminary ruling in Bartz that addressed the claim of copyright infringement and the defense of fair use. Throughout the Court’s consideration of the fair use factors, its analysis was quite divided depending on whether the works at hand were lawfully obtained or sourced from pirated repositories. With respect to the first fair use factor—the purpose and character of use—the Court found that Anthropic’s use of the authors’ copyrighted works to train their LLMs was “exceedingly transformative” and constituted fair use. (9) The Court’s ruling in favor of the transformative nature of Anthropic’s use is particularly significant to the broader fair use analysis. The characterization of a use as transformative carries substantial weight in copyright cases, as it distinguishes a secondary use that merely supersedes the original work from one that serves a new, distinct purpose. Regarding the second factor—the nature of the copyrighted work—Judge Alsup determined that Anthropic had chosen the authors’ works particularly due to their expressive qualities. This distinction is significant because works that are more creative or expressive in nature tend to receive greater copyright protection and copyists have less of a fair use defense than works that are primarily factual. Thus, the second factor weighed against fair use, as the expressive nature of the authors’ works enhanced their claim to copyright protection and diminished the availability of fair use as a defense. Interestingly, with regard to the third fair use factor—the amount and substantiality of the portion of the copyrighted work used in relation to its entirety—the Court explained that using any single work for training LLMs was “...about as reasonable as the next.” (8) Although the Court decided that the quantity of copied works was reasonable and necessary under the circumstances, it nevertheless found that the third factor weighed against fair use for works used in training that were pirated copies obtained illegally through shadow libraries. In contrast, however, the third fair use factor favored fair use for legally obtained works as they merely converted copies from print to digital. Finally, the fourth factor—the effect of the use upon the potential market—weighed in favor of fair use, though only as applied to the legally obtained copies. As for the legally purchased works, the Court rejected authors’ claims that Anthropic’s LLMs would produce “an explosion of works competing with their own,” reasoning that such an outcome is analogous to using copyrighted works to teach students to write well, which would similarly result in an increase in competing works. (9) Nonetheless, the use of pirated copies weighed against fair use under the fourth factor because the Court recognized that a market for licensing such works could develop, which Anthropic’s unauthorized use would effectively displace.
Following this preliminary summary judgment ruling, the case remained unresolved on several key issues, especially with regard to the use of pirated works. Because the Court had not yet issued a final determination, the case was set for trial. 17 U.S.C. § 504(c) describes the statutory damages for copyright infringement, stating that the plaintiff would be awarded between $750 and $30,000 if the defendant were found liable for infringement. Additionally, if the infringement was committed willfully, the plaintiff would be awarded up to $150,000. Given the large number of pirated works included in the eligible class, the damages could have reached as high as $14.5 billion, and even as much as $72.4 billion if the infringement was found to be willful. (10) However, rather than risk such uncertain liability, both parties ultimately chose to resolve the dispute through settlement.
The nature of Bartz as both a copyright infringement and class action lawsuit has resulted in a uniquely large gross settlement fund of $1.5 billion plus interest. Both aspects of this lawsuit, in combination, significantly increased Anthropic’s potential liability. Copyright law allows for statutory damages for each infringed work, while class certification allows claims involving hundreds of thousands of works to be consolidated into a single filing. As a result, even damages that are assessed at the lower end of the statutory range could collectively amount to billions of dollars across the sizable certified class. Due to the extraordinary size of this settlement, the Court reduced the attorneys’ fees from the typical 25-35% to 12.5%, amounting to $187.5 million, which is still quite a lucrative result for the attorneys representing the plaintiffs. In fact, the real recovery for each plaintiff is significantly smaller than what the lawyers were awarded. Thus, after all fees and litigation expenses, the estimated payout is about $3,000 per work, split between authors, co-authors, and publishers. (11) It is also especially important to note that as of April 2026, authors representing 440,490 of the 482,460 eligible works had chosen to join the certified class, resulting in a 91.3% claim rate, significantly greater than the average rate of about 10% in a typical class action lawsuit. This exceptional rate of participation is indicative of the sheer scale of alleged harm and the legitimacy of the claims against Anthropic. From a broader perspective, Bartz v. Anthropic highlights the immense legal and financial stakes that arise from the intersection of generative AI models and copyright law. As AI continues to advance, its training libraries will only grow larger, thus depending on increasing amounts of existing works. Courts will be forced to determine where to draw the line between transformative development and infringement of creators’ rights. How courts choose to strike this balance will determine the future for both the rights of copyright holders and possibly even how generative AI models are trained in the future.
Endnotes
1. Sara Merken, “Trouble with AI ‘Hallucinations’ Spreads to Big Law Firms,” Reuters,
May 23, 2025, https://www.reuters.com/legal/government/trouble-with-ai-hallucinations-spreads-big-law-firms-2025-05-23/.
2. Kate Knibbs, “Anthropic Settles High-Profile AI Copyright Lawsuit Brought by Book
Authors,” WIRED, August 26, 2025, https://www.wired.com/story/anthropic-settles-copyright-lawsuit-authors/.
3. “Bartz v. Anthropic,” BakerHostetler, accessed April 30, 2026,
https://www.bakerlaw.com/bartz-v-anthropic/.
4. Dave Hansen, “Bartz v Anthropic Settlement Gets Preliminary Approval – Key
Takeaways,” Authors Alliance, September 28, 2025, https://www.authorsalliance.org/2025/09/28/bartz-v-anthropic-settlement-gets-preliminary-approval-key-takeaways/.
5. Federal Rules of Civil Procedure 23,
6. U.S. Copyright Office, “Chapter 1: Subject Matter and Scope of Copyright,” Copyright
Law of the United States, accessed April 30, 2026, https://www.copyright.gov/title17/92chap1.html#107.
7. Rich Stim, “Measuring Fair Use: The Four Factors,” Stanford Copyright and Fair Use
Center, updated November 25, 2021, https://fairuse.stanford.edu/overview/fair-use/four-factors/.
8. Bartz v. Anthropic PBC, No. C 24-05417 WHA, Order on Fair Use (N.D. Cal. June 23,
2025), https://copyrightalliance.org/wp-content/uploads/2025/06/Bartz-v.-Anthropic-Order.pdf.
9. “Bartz v. Anthropic: First Court Decision on Fair Use Defense in LLM Training,” Wiggin
and Dana LLP, June 30, 2025, https://www.wiggin.com/publication/bartz-v-anthropic-first-court-decision-on-fair-use-defense-in-llm-training/.
10. 17 U.S.C. § 504,
https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title17-section504&num=0&edition=prelim.
11. “Anthropic Settlement Update: 91.3 Percent of Books Claimed in Settlement,” The
Authors Guild, April 17, 2026, https://authorsguild.org/news/anthropic-settlement-update-91-percent-of-books-claimed/.
