Recent U.S. court decisions have seemingly addressed whether the use of copyrighted materials for Artificial Intelligence (AI) training should be considered fair use. However, these decisions do not have the sweeping implications that the headlines might imply they do. In fact, these decisions are actually quite narrow and, to some extent, limited to the circumstances of the particular cases. Thus, based on the scope of these decisions, the open question we discuss here is: What do these decisions mean for those who wish to train AI systems with copyrighted material?
As a brief refresher, “fair use” is an affirmative defense to copyright infringement that, when successful, allows for the use of copyright-protected works by someone who is not the copyright holder, in which the use by the non-copyright holder would otherwise be considered infringement. To determine whether fair use applies, U.S. Copyright Law directs courts to consider a non-exhaustive four-factor list. The four factors are: (1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; (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; and (4) the effect of the use upon the potential market for or value of the copyrighted work.[1] While no factor is weighted more heavily than another, in many instances, the two most critical factors are the first, which determines the transformative nature of the use, and the fourth, which essentially determines the financial impact of the use.
In Bartz v. Anthropic PBC, Anthropic used books to train its large language model (LLM), which included both books that had been purchased by Anthropic and books that had not been purchased (i.e., pirated books). First, the court found that because the pirated books were impermissibly acquired, there was no fair use for the pirated books that Anthropic had used to train its LLM. For the purchased books, the court found that the use was transformative because the use of these works was to train the LLM, which would then generate new text and create something different. In relation to the fourth fair use factor, the court opined that this training of the LLM would not replace the demand for the authors’ works, and thus it had little impact on the market. The court reasoned that this was no different from training school children to write well.
In Kadrey v. Meta Platforms, Inc., Meta acquired books from shadow libraries, which provide works for free download, regardless of copyright status. The court found that because training an LLM is highly transformative, a fair use determination rested on the fourth fair use factor. The court specifically noted that the fair use ruling in favor of Meta was necessitated by the plaintiffs’ (the copyright holders) failure to present empirical evidence that Meta’s LLM outputs would harm the plaintiffs’ ability to profit from their own works.
The first distinctive characteristic about these cases is that they both involve the use of books for training an LLM. Thus, the applicability of these decisions and the fair use doctrine may vary for other types of works and other kinds of generative AI. For example, a fair use analysis for an image-generating AI would probably be very different. With text, which is used for training an LLM, such as in the Bartz and Kadrey cases, the LLM can learn the knowledge of the textual content and then, in responding to a prompt, generate a response using that knowledge without having to reproduce the original text. Additionally, detecting reproduced text is relatively straightforward and not subjective. In other words, not only is copying/reproduction obvious, but making even the slightest changes to the wording of text has the potential to be considered transformative for the purposes of fair use.
By contrast, determining fair use for images is already difficult when AI is not involved, as seen in the recent decision of Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith.[2] In that case, the artist Andy Warhol had taken a photograph of the musician Prince by Lynn Goldsmith and created a series of silkscreen illustrations. The Supreme Court found that Warhol’s changes were not sufficiently transformative for the use by Warhol to be considered fair use. It is objectively easy to determine if text has been copied (i.e., is transformative or not), but when the question of whether an artistic image is transformative has to go all the way to the Supreme Court for one of the most recognized artists of modern times, how will courts be able to make similar conclusions for the millions of images generated by AI? For example, how much does an AI image generator need to alter the likeness of Mickey Mouse for the output to be transformative and possibly considered fair use?
Another distinction for these cases is that the decisions are each very limited in some respects. For example, in Bartz, the purchased works for which fair use was found to be applicable would be destroyed once digitized and training was limited to these specific LLMs that the court found were transformative. For Bartz, at least, this would require the purchase of works that are being used for training the LLM with the assumption that the LLM would not actually reproduce the works. Leaning on the analogy the Bartz court made to teaching children to write, the underlying belief appears to be that the information of written works can always be reproduced by those who have consumed it, whether children, another person, or AI, but as long as the reproduction is not a direct copy, then it will never be as good as the original.
Kadrey is more lenient than Bartz regarding whether the works used for training need to be purchased. However, Kadrey is also very limited, because the court was unequivocal in the opinion that the ruling applied to only these thirteen plaintiffs for failing to present the correct arguments; this was not a class action, and the ruling does not mean that it is lawful for Meta to use any copyrighted works to train their LLM. So while a headline that read “Court Finds Training Meta’s LLM with Copyrighted Works is Fair Use” would be technically true, it is far from the whole truth. This decision, as the court stated, arises solely from the plaintiffs’ failure to present the proper arguments and has applicability limited to these thirteen plaintiffs.
These decisions demonstrate that the courts do not want to stymy the progress of AI and are open to the idea that training of AI using copyrighted works may be considered fair use, even by large corporations such as Meta. However, it is also clear that appropriate steps must be taken to avoid copyright infringement and/or meet fair use requirements. In Campbell v. Acuff-Rose Music, Inc., the Supreme Court held that the more transformative the use, the more heavily the first fair use factor will be weighed against the other three.[3] For both Bartz and Kadrey, the courts considered the training of the LLMs as a highly transformative use. The transformative nature of the use will continue to be the key factor for training AI.
Thus, while these decisions offer little insight into how fair use may be applied to other types of generative AI, they do provide some guidance for LLM developers/owners. In particular, these two decisions indicate that LLM developers/owners should consider attaining the works used for training legally, be able to demonstrate the transformative nature of the use (or at least show that their LLM does not produce copies), and try to identify evidence that the use (i.e. the training) of the copyrighted works will not have an effect on the market for the copyrighted works.
[1] See U.S. Copyright Office Fair Use Index; https://www.copyright.gov/fair-use/
[2] Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023).
[3] Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 584 (1994).
