Glaring Omissions from the Kadrey v. Meta Hearing
Last week, a much-anticipated hearing on motions for summary judgment was held in an ongoing infringement lawsuit brought by a group of authors against Meta for the unauthorized use of their literary works to train Meta’s Llama large language model. The hearing, which was overseen by Judge Vince Chhabria in the Northern District of California, was one of the first of its kind to focus on what has become the pivotal question in over 40 infringement lawsuits brought by copyright owners against AI companies: whether the unauthorized use of copyrighted works for training qualifies as fair use.
Judge Chhabria seemed genuinely skeptical of whether training qualifies as fair use as a general matter, but was clear that, in the case before him, he believes Meta’s use is transformative and the plaintiff authors have failed to show how Meta’s use of their works is harming (or has the potential) to harm the market for those works. Unfortunately, the hearing included little or no discussion of three key points that must be considered as part of any AI training fair use analysis: (1) under the first factor, whether the use was commercial or transformative and the ultimate purpose of the output, (2) the effect of Warhol v. Goldsmith on transformativeness and justification, and (3) the robust and growing market for AI licensing.
Commerciality and the Ultimate Purpose of Generative AI Training Not Addressed
One of the most surprising and unfortunate things about the nearly three-hour long hearing was the lack of any debate or substantive discussion over whether the purpose of Meta’s use of the plaintiffs’ literary works is actually transformative. At the start of the hearing, Judge Chhabria said that assuming a use is transformative, which in this case he said that it is, the most important question is whether there is market harm under the fourth fair use factor. And that was the extent of the transformative discussion, or for that matter any other discussion of the first fair use factor. Besides a few arguments made by Meta’s lawyer about how the more transformative a use, the less there can be a cognizable market effect, it never came up again. Shockingly, in the second half of the hearing, Kadrey’s lawyer did not address the issue or attempt to convince Chhabria that the purpose of Meta’s use was not transformative. Nor was there any discussion of the commerciality of Meta’s use and how that must be balanced against any a transformative use.
It was a major missed opportunity, because even though Judge Chhabria seemed set in his belief that Meta’s use was transformative, there is a very strong argument that, when considering the ultimate purpose of the use, it is not. As we explain in the amicus brief we filed in opposition of Meta’s motion for summary judgment, Meta and other proponents of AI training as fair use seek to isolate the training process and ignore the output of generative AI when describing the purpose of generative AI. In describing its purpose as “training,” Meta claims its copying “serves a manifestly different purpose from Plaintiffs’ books.” But Meta (and apparently Judge Chhabria) ignores what comes after the initial “training”—most notably the generation of output that serves the same purpose of the ingested works.
It should have been a point of discussion at the hearing that many of the past fair use cases Meta relies upon were careful to consider the ultimate purpose of a use and not simply end their analysis at an intermediate step—whether that be reverse engineering, shrinking a work into a thumbnail image for search engine purposes, or AI training. During the hearing, Judge Chhabria was right to distinguish the past fair use cases from generative AI training. But a principle from those cases that should have been recognized is that the purpose of any act of copying cannot be considered in a vacuum. Meta’s copying of the plaintiffs works for training must be considered alongside the ultimate purpose of generating material that undoubtedly serves the same purpose of the ingested works.
Judge Chhabria also seemed to assume that the fact that Meta’s use was transformative made the commerciality of its use irrelevant, since commerciality was never raised. As discussed more in the next section, even if the use is transformative, that in itself does not control the outcome of the first fair use factor, especially where the use is a commercial one, as it is here.
Discussion of Warhol’s Impact on Fair Use Analyses was Glaringly Absent
Another critical issue that was notably absent from the discussion is the Supreme Court’s clarification of the impact of a finding of transformativeness on a fair use analysis in Warhol v. Goldsmith. In Warhol, the Supreme Court confirmed that even when a court finds a use to be transformative, that finding should not only have a limited effect on the ultimate fair use determination, it also should not control a factor one analysis. Further, the Court explained that even when considering the purpose of a use, the defendant’s purported “justification” for the use must be considered as part of the first-factor analysis.
In reaffirming these crucial principles, the Warhol decision swung the fair use pendulum back to its intended position, and yet neither Judge Chhabria nor the parties’ lawyers raised “justification” during the hearing. While there was mention of the Warhol case by Meta’s attorney in an attempt to support the aforementioned argument that the more transformative a work, the less there is a market effect, Kadrey’s lawyer only raised the case once at the end of the hearing to make a tangential point about focusing the transformative analysis on a specific use. It’s hard to believe that such a seminal fair use case was barely discussed during a hearing on the fair use of AI ingestion. The Warhol case should have been the main event in the plaintiffs’ lawyer’s arguments and Judge Chhabria’s questions, but instead it was reduced to an afterthought throughout the entirety of the hearing.
Applying the Warhol standard, even if a use is found to be transformative, that must be weighed against other factor-one considerations like the above-mentioned “justification” for the use and the commercial nature of the use. Here, Meta’s use is clearly part of a commercial AI venture, designed to attract as many users as possible and bolster Meta’s position in the market. This commercial purpose should weigh against fair use under the first factor and offset in whole or in part any finding of transformative purpose. Hopefully Judge Chhabria will eventually consider commerciality in a written decision, but during the hearing it wasn’t even discussed.
Additionally, there was no discussion of the sufficiency or lack thereof of Meta’s justification for needing to copy the entire internet—including what its employees knew to be pirate websites—or ignoring existing licenses offered by copyright owners in order to train its generative AI models. AI companies may claim that they must copy all the copyrighted material they can get their hands on to train an optimally functioning model, but the truth is that there are plenty of successful models that have been developed using proprietary or licensed materials and materials in the public domain. It will be interesting to see if or how Judge Chhabria handles the “justification” issue in a written decision because if he ignores it there too, it will likely be the first issue raised on appeal.
The important takeaway here is that even if the use of copyrighted works for ingestion by Meta is found to be a transformative use in this case, that would not necessarily mean factor one favors fair use as Judge Chhabria seemed to assume. It is essential that plaintiffs in AI infringement cases understand and argue this fundamental tenet that Warhol confirmed, and judges overseeing the cases must recognize its impact on fair use analyses.
AI Licensing Markets Must be Considered Under Factor Four
During the hearing, Judge Chhabria said that he would not take into account AI licensing markets when considering market harm under the fourth factor, indicating that AI licensing is too “circular.” What he meant is that if AI training qualifies as fair use, then there is no need to license and therefore no harmful market effect. But that reasoning ignores the important realities of an already robust market for licensing of AI training material that must be considered under factor four—realities that should have been raised by Kadrey’s lawyer and recognized by courts moving forward. The licensing argument could only be circular if it’s a potential licensing market, not if it’s an actual market that already exists. Since a robust market does exist, Judge Chhabria and Meta’s counsel are incorrect to summarily claim the argument is circular.
In addition, circularity could be argued in nearly any fair use case in which licensing would become obsolete or less necessary if the use was found to be fair use. But courts routinely consider the impact an unauthorized use would have on licensing markets if the use were to become widespread, regardless of an eventual finding of fair use. For example, in Campbell v. Acuff Rose, the Supreme Court explained that “[t]he market for potential derivative uses includes only those that creators of original works would in general develop or license others to develop.” While AI training is not the type of derivative use the Supreme Court had in mind in Campbell, the logic can be applied just the same. If the market for licensing to others for AI training uses were deemed inapplicable under a fourth factor analysis, the same could be done for markets for any derivative uses that an alleged infringer claims qualify as fair use. And that would effectively rob copyright owners of their guaranteed rights.
In the case of generative AI training, the licensing markets are not hypothetical; they exist and continue to grow as AI companies realize the value in training on high quality, curated datasets of copyrighted materials. In a brief filed in opposition of Meta, the Association of American Publishers goes into great detail on the robust licensing market that already exists and continues to grow for AI training. These deals are clear evidence that AI companies recognize the value of the copyrighted works they want for training and disrupting that licensing market would undoubtedly threaten the ecosystems that enable the creation of human-authored works that generative AI will need in the future.
Conclusion
The Kadrey v. Meta hearing failed to address critical fair use principles that are now undeniable after Warhol. While it’s unfortunate that these issues we’re left unaddressed, the hope is that Meta’s motion for summary judgment is denied, and these issues can be addressed at trial.
If you aren’t already a member of the Copyright Alliance, you can join today by completing our Individual Creator Members membership form! Members gain access to monthly newsletters, educational webinars, and so much more — all for free!