Fair Use Week is an annual celebration of the legal doctrine of fair use, which plays an essential role in teaching, education, and scholarship. The fair use doctrine allows for the use of copyrighted works in certain circumstances, which is determined using a four-factor test that considers the purpose of the use, the nature of the copyrighted work, the amount and substantiality used, and the effect of the use on the market for the copyrighted work.
Fair Use in Current Artificial Intelligence Lawsuits
A year ago, in recognition of Fair Use Week, Copyright Corner wrote about the intersection of fair use and artificial intelligence (AI). It explored how AI might interact with a four-factor fair use analysis, summarized guidance from Parts 1 and 2 of the U.S. Copyright Office’s Copyright and Artificial Intelligence report, and provided an overview of some current AI lawsuits involving copyright. At the time, a new summary judgement ruling in Thomson Reuters v Ross Intelligence had been recently released, giving an early look into how courts might rule on fair use in the context of AI. Since then, Part 3: Generative AI Training of the U.S. Copyright Office’s AI report was published, and further rulings in current cases involving copyright and AI have been released, which both expand on how the standing legal frameworks in copyright law, and more specifically, fair use, may be applied to the unprecedented advances in AI technology. To keep up with how courts are applying established principles of the fair use doctrine to AI, let’s explore some of the key takeaways from recent decisions in Bartz v. Anthropic and Kadrey v. Meta.
In Reuters v. Ross, the court rejected a fair use defense in the use of copyrighted works for training of an AI legal search tool. Weighing the first factor, the court held that Ross was using the headnotes as AI data to create a competing legal research tool, which was not a transformative use. Additionally, under the fourth factor, the court found that Ross’s legal research tool served as a market substitute and also noted consideration for the effect of Ross’s use on a potential market for AI training data. Perhaps unsurprising to those familiar with fair use analyses, those same two factors—the first and the fourth—play a large role in the decisions in Bartz and Kadrey, which were released in the same week in June 2025, only two days apart. In both cases, lawsuits were brought against companies that used copyrighted works to train large language models without authorization from the copyright owners. Both defendants made fair use defenses, and in both cases, the judges ruled in favor of fair use, at least in part. There is much to be gathered, however, beyond a simple fair use finding from the analysis the judges provide in their opinions.
In Bartz v. Anthropic, while the court denied summary judgment for Anthropic’s use of pirated copies of copyrighted works to assemble a central library, it also found that training AI on copyrighted works is a fair use if the works are acquired legally. According to the judge, Anthropic’s use of legally acquired works to train AI models was “transformative—spectacularly so,”[1] which supports a fair use finding under the first factor. When evaluating any potential market harm under the fourth factor, the judge found “copies used to train specific LLMs did not and will not displace demand for copies of Author’s works” because there is no evidence that the AI would produce exact copies of the works, meaning there is no direct substitution.[2] The judge went on to analogize the plaintiff’s argument for potential market harm: “Authors’ complaint is no different than it would be if they complained that training school children to write well would result in an explosion of competing works.”[3] Ultimately, the decision in Bartz v. Anthropic suggests that companies training AI on copyrighted works may be able to rely on fair use when those works they are using are lawfully acquired. For the issue of Anthropic’s use of pirated copies of works, notable developments in the case came in August and September 2025 when it was certified as a class action and preliminary approval of a $1.5 billion settlement was announced. Nearly half a million works have claims in this case through the certified class, meaning the potential statutory damages Anthropic could have faced had they lost were astronomical, so they decided to settle for a smaller amount to avert that risk.
The judge’s opinion in Kadrey v. Meta, despite an overall finding of fair use, proved to be more controversial. Similar to Bartz, the judge in Kadrey found the use of copyrighted works as training data for AI models to be transformative, noting that the original purpose of the works was to be “read for entertainment or education” and using the works in various AI functions was a sufficiently different purpose.[4] Approaching the fourth factor, however, is where the Kadrey decisions differs significantly from the analysis in Bartz. The judge in Kadrey employs a novel “market dilution” theory, arguing that “[n]o other use…has anything near the potential to flood the market with competing works the way that LLM training does.”[5] In the judge’s opinion, the competing works serve as indirect substitutes for the copyrighted works (instead of the typical direct substitutes that are considered under the fourth factor in a fair use analysis). The judge finds that even a transformative use can produce an indirect market substitute under the market dilution theory: “No matter how transformative LLM may be, it’s hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works that could significantly harm the market for those books.”[6] The same analogy from Bartz about teaching children to write is addressed in Kadrey, with the judge arguing that “using books to teach children to write is not remotely like using books to create a product that a single individual could employ to generate countless competing works with a miniscule fraction of the time and creativity it would otherwise take.”[7] Ultimately, the opinion in Kadrey emphasizes the unprecedented nature of AI technology and how it may affect a copyright owner’s ability capitalize on their work. Though the court ultimately grants summary judgement for Meta on its fair use defense,[8] it outlines an analysis that may prove prohibitive for those relying on fair use to train AI models on copyrighted works.
It is important to remember that each fair use case is different and highly fact dependent, so results of fair use and AI cases will likely continue to vary based on the facts of a particular case, just as Bartz and Kadrey did, so it may be wise to hesitate from drawing generalizations from the results of only two cases. Hopefully, as courts perform more fair use analyses and subsequently release more opinions, a more clear and consistent thread of practices will become evident, giving a clearer picture on what is and is not fair use in the context of AI.
[1] Bartz v. Anthropic PBC, 3:24-cv-05417, (N.D. Cal.). gov.uscourts.cand.434709.231.0.pdf
[2] Id.
[3] Id.
[4] Kadrey v. Meta Platforms, Inc., 3:23-cv-03417, (N.D. Cal.) gov.uscourts.cand.415175.598.0.pdf
[5] Id.
[6] Id.
[7] Id.
[8] According to the judge in Kadrey, “the plaintiffs presented no meaningful evidence on market dilution at all,” which was a driving factor in the fourth factor favoring the defendant and resulted in a finding of fair use. Had the plaintiffs been able to demonstrate market dilution, the fourth factor may have gone to a jury or they may have been able “to win on the fair use issue at summary judgment.”
By Landen Stafford (Copyright Services Specialist at Copyright Services, The Ohio State University Libraries)
Recent Comments