Introduction
Recent rulings by two judges have addressed the issue of fair use in the context of AI model training by companies Anthropic and Meta. These decisions mark significant developments in the legal landscape surrounding generative AI systems [1], though neither ruling represents a definitive victory for the involved companies.
Description
Two judges recently issued rulings regarding fair use in the context of AI model training by Anthropic and Meta [4], marking significant developments in the legal landscape surrounding generative AI systems [1]. However, neither decision can be considered a clear victory for the companies involved [4]. In the case of Bartz v. Anthropic [1] [2] [3] [4] [5] [6] [7] [8], US District Judge William Alsup determined that Anthropic’s downloading of over seven million “pirated books” from online shadow libraries constituted infringement [4]. He emphasized that the method of obtaining these books is crucial [6], distinguishing between lawful purchases and the unacceptable act of downloading pirated copies [6]. This led to a separate trial focused on potential damages related to these unauthorized copies, as the court noted that Anthropic opted to “steal” books instead of pursuing licensing agreements [2]. While the judge recognized the transformative nature of Anthropic’s use of copyrighted material for training its Claude large language models (LLMs), describing it as “exceedingly transformative,” he expressed skepticism about justifications for using pirated materials, stating that purchasing a book after stealing it does not absolve liability [6].
Anthropic expressed satisfaction with the ruling on transformative use but disagreed with the decision regarding the downloading of books and is considering its options for appeal [1]. The case was initiated by a group of authors [1], and their legal team has not yet indicated whether they will pursue an appeal [1]. Content creator groups are encouraged by the possibility that Anthropic could be held accountable for its method of accessing copyrighted material [1], which may prompt AI companies to seek permission from copyright holders for training their models [1].
Conversely [4], Judge Chhabria ruled in favor of Meta regarding its use of copyrighted materials for AI training [8], although he remains unconvinced that this use qualifies as fair use [8]. He found that the authors did not provide sufficient evidence to demonstrate market harm under a new theory of copyright dilution [4], although he indicated that most uses of copyrighted works for AI training without permission are likely illegal [4]. In Meta’s case, the downloading of copyrighted material was considered part of the same use as training its AI model, which was ultimately found to be fair use. Judge Chhabria dismissed two key arguments from the authors [8], stating that Meta’s Llama AI does not reproduce enough text from their works to be significant and that the authors are not entitled to license their works specifically for AI training purposes [8]. Meta expressed support for the rulings [3], emphasizing the role of open-source AI models in driving innovation [3]. However, ongoing lawsuits against AI companies [5], including a separate case against Meta [5], highlight the potential for substantial liability for using pirated materials, which could result in significant statutory damages [5].
The first factor of fair use [4], concerning the purpose and character of the use [4], favored fair use in both cases [4]. Anthropic’s training of LLMs was seen as transformative, producing new [4], non-infringing outputs [4] [5]. Similarly [4], Meta’s use was characterized as highly transformative [4], aimed at developing innovative tools that generate diverse text.
Regarding the nature of the copyrighted work [4], both rulings disfavored fair use due to the creative nature of the books involved [4], although the transformative purpose lessened the weight of this factor [4]. The amount and substantiality of the portion used favored fair use in both cases [4], as the benefits of training LLMs outweighed any public revelations of the works themselves [4].
The fourth factor [4], concerning the effect on the potential market for the copyrighted work [4], also favored fair use for both companies [4]. Anthropic’s use did not result in market harm [4], while Meta’s use was seen as potentially beneficial [4]. However, Judge Chhabria accepted the new theory of market dilution [4], suggesting that training LLMs with copyrighted books could harm the market for those works [4], even if the generated outputs were not infringing [4]. This raises concerns about ongoing lawsuits from authors, publishers [3], and artists [3] [5], particularly in sectors like news publishing [3], which may have stronger claims against AI tools [3].
Overall, while both decisions found fair use [4], they included significant caveats [4]. Judge Alsup’s ruling indicated that acquiring pirated copies would likely be treated as infringement [4], while Judge Chhabria’s acceptance of market dilution raises important questions about the legality of AI training practices involving copyrighted materials [4]. The potential penalties for copyright infringement in such cases could amount to billions [7], given the scale of the pirated library involved in Anthropic’s case [7]. The rulings serve as a significant precedent for AI companies [5], indicating that while training can be considered fair use [5], there are important qualifications that must be navigated carefully. Courts are expected to continue shaping the legal framework as technology evolves [1], with no clear legislative direction from Congress on this issue [1].
Conclusion
These rulings underscore the complexity of applying fair use doctrine to AI training, highlighting the need for AI companies to carefully navigate copyright laws. While both Anthropic and Meta received favorable rulings on transformative use, the decisions also emphasize the potential liabilities associated with using pirated materials. The evolving legal landscape suggests that AI companies must be vigilant in securing permissions and licenses to mitigate risks, as courts continue to refine the legal framework in response to technological advancements.
References
[1] https://deadline.com/2025/06/anthropic-fair-use-copyright-1236442323/
[2] https://www.publishersweekly.com/pw/by-topic/digital/copyright/article/98089-federal-judge-rules-ai-training-is-fair-use-in-anthropic-copyright-case.html
[3] https://observer.com/2025/06/meta-anthropic-fair-use-wins-ai-copyright-cases/
[4] https://chatgptiseatingtheworld.com/2025/06/26/12026/
[5] https://news.bloomberglaw.com/ip-law/mixed-anthropic-ruling-builds-roadmap-for-generative-ai-fair-use
[6] https://fortune.com/2025/06/24/ai-training-is-fair-use-federal-judge-rules-anthropic-copyright-case/
[7] https://www.wired.com/story/anthropic-ai-copyright-fair-use-piracy-ruling/
[8] https://www.theverge.com/news/693437/meta-ai-copyright-win-fair-use-warning