September 2025
The Boundaries of Fair Use in AI Training: Insights from Kadrey et al. v. Meta Platforms, Inc. in the U.S.
In June 2025, the United States District Court for the Northern District of California issued a ruling on Meta’s motion for summary judgment in a copyright suit brought by 13 authors
[1]
. The dispute concerned whether using copyrighted books to train generative AI models constitutes fair use under U.S. copyright law. This decision, following closely after the court’s ruling in the Anthropic case
[2]
, once again addressed the boundaries of fair use in the context of AI training.
While both Meta and Anthropic were accused of downloading copyrighted books from the internet to train their AI models, the judges in each case emphasized different fair use factors, leading to divergent conclusions. In the Anthropic case, the judge found that AI training was sufficiently transformative to qualify as fair use. By contrast, the judge in the Meta case underscored that the purpose of copyright law is to protect authors’ incentives to create, thereby encouraging further creation across society. Therefore, although the use of copyrighted books for AI training may be transformative, if such use causes substantial impairment of the potential market for the original works, it does not qualify as fair use.
1. Background
In addition to social media platforms such as Facebook and Instagram, Meta also offers a suite of large language models called “Llama”. By 2025, Llama was projected to generate between USD 2 billion and 3 billion in revenue for Meta.
To train Llama, Meta obtained vast quantities of books from “shadow libraries”—websites providing unauthorized access to academic articles and books—including books authored by the 13 plaintiffs in this case (“Plaintiffs’ books”).
The plaintiffs sued Meta, alleging that using their books without authorization to train Llama constituted copyright infringement. They argued that Llama could generate text containing content from their books and that Meta’s actions impaired the potential market for licensing books specifically for AI training.
In response, Meta argued that its use of Plaintiffs’ books constituted fair use. Both parties filed motions for summary judgment, asking the court to rule on the issue of fair use.
2. Court’s Analysis
(1) The Use of Plaintiffs’ Books for AI Training Is Highly Transformative
The court held that Meta used Plaintiffs’ books to train Llama for the purpose of generating new works different from the Plaintiffs’ books, which are meant to be read and enjoyed. Although Llama might generate content similar in style, copyright law protects the expression itself rather than the style, and therefore the purposes should not be regarded as the same. For these reasons, the court determined that Meta’s use of Plaintiffs’ books was highly transformative, consistent with the conclusion reached in the Anthropic case [3] .
(2) Downloading and Training Should Be Evaluated Together
The court further held that Meta’s downloading of pirated copies of books from shadow libraries should not be evaluated separately from its subsequent use in AI training. The downloading was merely instrumental to the training. Since the use of Plaintiffs’ books for AI training was transformative, the downloading was also deemed transformative. This reasoning diverged from the Anthropic case, where the judge considered the downloading of unauthorized copies of books itself to be an act of infringement [4] .
(3) Transformative Use Is Not the Only Consideration: Potential Market Impairment Should Also Be Evaluated
The court emphasized that copyright law exists to protect authors’ incentives to create. Thus, even Meta’s transformative use may not qualify as fair use if it causes substantial impairment of the potential market for Plaintiffs’ books. This point was not addressed in the Anthropic case [5] .
The court noted that unauthorized use of Plaintiffs’ books for AI training could result in “market dilution”, where Llama can generate vast quantities of content similar in subject or style at minimal cost and time. Such content could subject Plaintiffs’ books to overwhelming competition, potentially causing them to be drowned out or substituted, thereby seriously undermining authors’ incentives to create.
The court elaborated on specific considerations in assessing market dilution. For example, “functional works” (e.g., news articles) are more vulnerable to dilution than “creative works” (e.g., novels), since readers often consume functional works primarily for information regardless of authorship. Similarly, books by lesser-known authors are more susceptible to dilution than those by highly renowned authors.
In this case, however, the plaintiffs failed to provide sufficient evidence of market dilution, and the court accordingly granted summary judgment in Meta’s favor.
3. Practical Insights
The Meta and Anthropic rulings are both landmark cases concerning generative AI training, yet they reflect divergent judicial reasoning. In the Meta case, the court declined to treat transformative use as dispositive. Instead, it underscored the principle of protecting authors’ incentives to create, placing potential market impairment as the central consideration in the fair use analysis. If substantial impairment is found, the use does not qualify as fair use.
It is important to note that the plaintiffs did not prevail due to insufficient evidence, not because the court wholly endorsed Meta’s use of Plaintiffs’ books. Thus, the ruling should not be seen as a blanket approval of generative AI companies’ use of copyrighted works. Compared to the Anthropic case, the Meta ruling reflects a stricter judicial approach toward the use of copyrighted books for AI training, which generative AI companies should approach with caution.
Given that U.S. judicial practice is highly likely to exert significant influence on future court decisions in Taiwan, LTP will continue to closely monitor these judicial developments.
[1] Kadrey, et al. v. Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal.).
[2] Jane Tsai & Vincent Kuo (2025), The Boundaries of Fair Use in AI Training: Insights from Bartz et al. v. Anthropic PBC in the U.S., Lee, Tsai & Partners.
[3] Id.
[4] Id.
[5] Id.
While both Meta and Anthropic were accused of downloading copyrighted books from the internet to train their AI models, the judges in each case emphasized different fair use factors, leading to divergent conclusions. In the Anthropic case, the judge found that AI training was sufficiently transformative to qualify as fair use. By contrast, the judge in the Meta case underscored that the purpose of copyright law is to protect authors’ incentives to create, thereby encouraging further creation across society. Therefore, although the use of copyrighted books for AI training may be transformative, if such use causes substantial impairment of the potential market for the original works, it does not qualify as fair use.
1. Background
In addition to social media platforms such as Facebook and Instagram, Meta also offers a suite of large language models called “Llama”. By 2025, Llama was projected to generate between USD 2 billion and 3 billion in revenue for Meta.
To train Llama, Meta obtained vast quantities of books from “shadow libraries”—websites providing unauthorized access to academic articles and books—including books authored by the 13 plaintiffs in this case (“Plaintiffs’ books”).
The plaintiffs sued Meta, alleging that using their books without authorization to train Llama constituted copyright infringement. They argued that Llama could generate text containing content from their books and that Meta’s actions impaired the potential market for licensing books specifically for AI training.
In response, Meta argued that its use of Plaintiffs’ books constituted fair use. Both parties filed motions for summary judgment, asking the court to rule on the issue of fair use.
2. Court’s Analysis
(1) The Use of Plaintiffs’ Books for AI Training Is Highly Transformative
The court held that Meta used Plaintiffs’ books to train Llama for the purpose of generating new works different from the Plaintiffs’ books, which are meant to be read and enjoyed. Although Llama might generate content similar in style, copyright law protects the expression itself rather than the style, and therefore the purposes should not be regarded as the same. For these reasons, the court determined that Meta’s use of Plaintiffs’ books was highly transformative, consistent with the conclusion reached in the Anthropic case [3] .
(2) Downloading and Training Should Be Evaluated Together
The court further held that Meta’s downloading of pirated copies of books from shadow libraries should not be evaluated separately from its subsequent use in AI training. The downloading was merely instrumental to the training. Since the use of Plaintiffs’ books for AI training was transformative, the downloading was also deemed transformative. This reasoning diverged from the Anthropic case, where the judge considered the downloading of unauthorized copies of books itself to be an act of infringement [4] .
(3) Transformative Use Is Not the Only Consideration: Potential Market Impairment Should Also Be Evaluated
The court emphasized that copyright law exists to protect authors’ incentives to create. Thus, even Meta’s transformative use may not qualify as fair use if it causes substantial impairment of the potential market for Plaintiffs’ books. This point was not addressed in the Anthropic case [5] .
The court noted that unauthorized use of Plaintiffs’ books for AI training could result in “market dilution”, where Llama can generate vast quantities of content similar in subject or style at minimal cost and time. Such content could subject Plaintiffs’ books to overwhelming competition, potentially causing them to be drowned out or substituted, thereby seriously undermining authors’ incentives to create.
The court elaborated on specific considerations in assessing market dilution. For example, “functional works” (e.g., news articles) are more vulnerable to dilution than “creative works” (e.g., novels), since readers often consume functional works primarily for information regardless of authorship. Similarly, books by lesser-known authors are more susceptible to dilution than those by highly renowned authors.
In this case, however, the plaintiffs failed to provide sufficient evidence of market dilution, and the court accordingly granted summary judgment in Meta’s favor.
3. Practical Insights
The Meta and Anthropic rulings are both landmark cases concerning generative AI training, yet they reflect divergent judicial reasoning. In the Meta case, the court declined to treat transformative use as dispositive. Instead, it underscored the principle of protecting authors’ incentives to create, placing potential market impairment as the central consideration in the fair use analysis. If substantial impairment is found, the use does not qualify as fair use.
It is important to note that the plaintiffs did not prevail due to insufficient evidence, not because the court wholly endorsed Meta’s use of Plaintiffs’ books. Thus, the ruling should not be seen as a blanket approval of generative AI companies’ use of copyrighted works. Compared to the Anthropic case, the Meta ruling reflects a stricter judicial approach toward the use of copyrighted books for AI training, which generative AI companies should approach with caution.
Given that U.S. judicial practice is highly likely to exert significant influence on future court decisions in Taiwan, LTP will continue to closely monitor these judicial developments.
[1] Kadrey, et al. v. Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal.).
[2] Jane Tsai & Vincent Kuo (2025), The Boundaries of Fair Use in AI Training: Insights from Bartz et al. v. Anthropic PBC in the U.S., Lee, Tsai & Partners.
[3] Id.
[4] Id.
[5] Id.
