United States - Richard Kadrey et al. v. Meta Platforms Inc. - (Artificial) market harm as a decisive factor

On 25 June 2025, the US District Court, Northern District of California issued a summary judgment in the case Richard Kadrey et al. v. Meta Platforms Inc., ruling that Meta Platforms Inc. (Meta) did not infringe upon the plaintiffs' copyrights by using their books to train its large language models (LLMs). This case is significant as it marks one of the first copyright infringement rulings related to the training of LLMs, following a similar decision in favor of the AI software company Anthropic PBC (Anthropic) on the training use.

Meta, known for its social media platforms, such as Facebook, Instagram and WhatsApp, is also the developer of Llama, a series of LLMs. To train these models, Meta initially sought to obtain books legally through licensing agreements with major publishers. However, after facing logistical challenges, the company began downloading books from “shadow libraries”, which provided access to copyrighted content without permission from the creator.

The plaintiffs are thirteen authors whose works are included in the datasets used to train Meta's LLMs. They have sued Meta for copyright infringement, among other claims. Before the case was fully resolved, both parties sought summary judgment on whether using copyrighted material to train Meta's LLMs without the creators' permission constitutes copyright infringement or if it is permissible under the fair use defense.

Judge Chhabria commenced with stating that the ultimate goal of copyrights is to promote a broad public availability of literature, music and other arts. Copyright law does this by granting artists incentives to create by giving them exclusive rights. An example is the right to prevent others from reproducing their works. However, copyrights law also seeks to strike a balance between protecting ownership and allowing room for innovation through the fair use doctrine. Fair use is defined in Section 107 of the Copyright Act, which allows copyrighted works to be used without the creator’s permission if certain criteria are met. The Copyright Act outlines four factors that should be considered when assessing whether a particular use is fair:

1. The purpose and character of the use;
2. The nature of the copyrighted work;
3. The amount and substantiality of the portion used in relation to the copyrighted work as a whole;
4. The effect of the use upon the potential market for or value of the copyrighted work.

Meta invoked this fair use doctrine. The first factor favors Meta, because the purpose of the thirteen authors and Meta clearly differ. There is no doubt that Meta’s use of the copyrighted works is highly transformative. In this case the use by Meta was to train its LLMs, which are innovative tools that can be used to generate a variety of texts and perform a wide range of functions. By contrast, the author’s books are intended for entertainment or education purposes.

With regard to the second factor the Court was brief. Meta chose these books precisely because of their strongly expressive value. Works that are more creative are more likely to receive greater copyright protection and are less likely to fall under the fair use doctrine. The second factor favors the thirteen authors, as there is consensus that LLMs perform better when trained on high-quality material involving multiple creative choices.

The third factor requires the amount and substantiality of the portion used in the new content to be reasonable in relation to the purpose of copying. This is related to the first factor. The extent to which new content may contain fragments of copyright-protected works is determined by the purpose of its use. Given the transformative purpose, judge Chhabria decided the amount copied by Meta is reasonable. Another relevant consideration is the amount of copyrighted material made available to the public, rather than the amount used by the alleged infringer. Experts from both parties reached the conclusion that Llama could not regurgitate more than fifty words from any of the authors’ books, even when adversarial prompting techniques were deployed.

The Court emphasizes that great importance must be given to the fourth factor, which considers the impact upon the potential market for or value of the copyrighted work. While Meta suggested that the transformative use of the copyrighted works constituted fair use, Judge Chhabria did not follow their reasoning. Even if the use is transformative, it can still have an impact on the market or on the value of the copyrighted work. The Court highlights a promising argument that provides support for this: an LLM can potentially generate works that compete with the original work by acting as indirect substitutes for it. Although the authors are unable to provide satisfactory evidence, the Court zooms in on this point. The training of an LLM can help to enable the rapid creation of countless works that can compete with copyrighted works, even if those works are not infringing any copyrights. This harm of market dilution – also referred to as harm of indirect substitution – will maybe not crowd out bestselling authors, but it is imaginable that lesser-known authors or up-and-coming authors will encounter difficulties due to artificial competition on the market. The plaintiffs did not present any empirical evidence that the training of LLMs is already harming the market or will harm it in the future, but instead solely offered speculation to the Court. Given the combination of the foregoing and Meta’s demonstration of a lack of market harm, the Court followed Meta on its fair use defense.

Judge Chhabria emphasized the limited consequences of this ruling, by noting the plaintiffs’ lack of evidence, which hindered their market dilution argument, and by pointing out the fact that this is not a class action, meaning the ruling only affects the rights of these thirteen authors.

He further stated that “it seems likely that market dilution will often cause plaintiffs to decisively win the fourth factor—and thus win the fair use question overall—in cases like this.” This gives creators whose works are being used to train LLMs without their permission a potentially strong argument for future lawsuits. Whether this argument will hold up in other cases, remains to be seen.

This case is significant as it marks one of the first copyright infringement rulings related to the training of LLMs, following a similar decision in favor of Anthropic. However, there are some differences between the Meta ruling and the Anthropic ruling. First, unlike the Anthropic case, Judge Chhabria placed greater emphasis on the impact on the potential market, arguing it may outweigh other considerations, including the transformative nature of the work. Second, while this ruling appears to favor Meta and other AI developers, it also equips authors and creators with a strong argument for future cases, potentially establishing that training generative AI on copyrighted works without permission is not permissible under fair use and constitutes copyright infringement.

Koen Vranckaert
Sultan Erdogan
Shannen Verlee