The First Major AI Fair Use Rulings: What the Anthropic and Meta Cases Mean
Unpacking Bartz et al v. Anthropic and how it fits into the broader context
Fair use is in the air of AI.1 It is fair to say that AI has been the most controversial topic in content for decades, at least since the rise of the internet. Content is used, in massive quantities, to train the models that give us so-called “foundational” AI models. AI can also be used to create content by lowering the cost of producing said content, especially when it is of the same form and style of the content it was trained on. Is this fair use, or copying? Is it a tool for creativity, or an infringement machine for “regurgitating” copies and creating derivative works, especially in light of recent precedents like Williams v. Gaye, No. 15-56880 (9th Cir. 2018) and Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023)? The answer is worth trillions of dollars, spawning a multitude of lawsuits in record time. Santa Clara University’s Edward Lee has been tracking these lawsuits on
, has been tracking these lawsuits; as of this writing, there are nearly 50 lawsuits, including notable ones like the New York Times lawsuit against OpenAI.Most of these suits are seeing their claims dismissed or settled, so there has not been much opportunity to test these legal theories in court. The first of these came down in February, Thomson Reuters Enter. Ctr. GmbH et al. v. ROSS Intelligence Inc., No. 1:20-cv-00613-SB, 2025 WL 458520 (D. Del. Feb. 11, 2025). This week, we got two decisions which may turn out to be seminal: Bartz v. Anthropic PBC, 3:24-cv-05417, (N.D. Cal. Jun. 23, 2025) and Kadrey v. Meta Platforms, Inc., 3:23-cv-03417, (N.D. Cal. Jun. 25, 2025). This week on Nonobvious, we will review these cases and what they mean from the perspective of fair use and AI.
Where We’re Going, We Don’t Need Unfair Use
Without the ability to train a model, there is no AI. Therefore, if rightsholders can prevent model providers from training on their data, they can prevent the creation of AI. Unlike a library, where you can remove the offending book, models “embed” data in a nonliteral way as part of a broader network of understanding; similar to how a human cannot unlearn something they’ve read, a model training run, which costs tens of millions of dollars, cannot be redone to remove the knowledge. The upshot is that a finding against fair use would result in an unworkable remedy or allow a rightsholder to remove their corpus from the body training, which would affect the quality of all models going forward. Add to the complexity that the scale of the corpus is more important than any particular work and this all gets complex fast.
As a reminder, there are four factors: (1) the character of the use, with a focus on whether the work is transformative and commercial; (2) the nature of the copyrighted work, with a balance in favor of fair use in factual contexts; (3) the amount amount and substantiality of the portion copied; and (4) the effect of the use on the potential market. While factor four is typically the most important, the others do still matter, especially the transformative nature of the work.2
AI is a particularly interesting case for fair use. It involves making literal copies of works and potentially transforming them into digitally useful forms; the works involve not just facts but also the most creative works, like poetry and music; there is debate on how transformative the models are, particularly since they can be tricked into “regurgitating” portions of complete works; and because of the potential impact on the market for the works themselves. Adding to the matter is that modern foundational models require truly gargantuan amounts of data, which is rumored to involve the use of well-known datasets with entries of, let’s say, sometimes dubious provenance.
Factually, the way the data was obtained is the main difference between the Anthropic and Meta cases. In the case of Anthropic, the company used a combination of admittedly pirated works obtained in a digitally native fashion—in this case, the subtly named Pirate Library Mirror—as well as purchasing over two million used books which it then destructively digitized. In the case of Meta, famous authors including Sarah Silverman objected to the mere inclusion of their book in the dataset, regardless of how it was obtained (though in the case of Meta, although it is alleged that they use the notorious pirated library LibGen, there was no discovery to prove it). Otherwise, the question of fair use is basically the same.
Digitization was only an issue in Anthropic, and Judge Alsup held those efforts to be allowed as fair use under precedents like Sony Computer Entertainment, Inc. v. Connectix Corp., 203 F.3d 596 (9th Cir. 2000) and Sega Enterprises Ltd. v. Accolade,
Inc., 977 F.2d 1510 (9th Cir. 1992), which focus on the “ultimate use” of the copying. Judge Alsup, in particular, focused on the lack of the “multiplication” problem present in A&M Records, Inc. v. Napster, Inc., 239 F.3d 1004 (9th Cir. 2001) because Napster distributed sometimes millions of copies of the works while Anthropic was merely changing the form of the works for its own use. When it comes to piracy, however, Judge Alsup clearly comes down hard against it. In Meta Platforms, Judge Chhabria surprisingly does not hold Meta liable for downloading “shadow” copies of works they have purchased separately.3
Both cases grant partial motions for summary judgement, finding that in these cases, training a large language model on a corpus is fair use—notably, this case is not about the outputs of the models, so derivative works were not at issue. A casual reader might thus conclude that these cases come to the same conclusion. But they are actually quite different. In Anthropic, Judge Alsup comes out swinging, arguing that model training is obviously transformative and writing that “to make anyone pay specifically for the use of a book each time they read it, each time they recall it from memory, each time they later draw upon it when writing new things in new ways would be unthinkable.” In Meta Platforms, in contrast, Judge Chhabria does say that “there’s no disputing that” large language models are transformative, he also says on page one of the opinion that “in most cases” such copying will be illegal, though he does not seem to have a deep theory for why this would be an exception. It appears that he relies almost entirely on the transformativeness factor even though a great deal of time is devoted to a market dilution theory of harm for the fourth factor. It feels similar to the famous Google Books case, Authors Guild v. Google 804 F.3d 202 (2nd Cir. 2015),4 where the use was so transformative, but also so useful, that it had to be allowed.
The other big distinction is on the effect on the market. In Anthropic, Judge Alsup focuses on the AI models’ inability to compete directly with their works. Instead, he almost derisively gestures towards the contentions of the authors that the “explosion” of competing works is a “generic” contention. He is surely right that it is speculative, but there is not a deeper analysis here, which leaves the reader wanting. In contrast, Judge Chhabria goes deep into this analysis, but the analysis itself falls short by failing to understand the zero marginal production cost of informational works. In one passage, he compares the effect of training an AI model on the life of President Lyndon B. Johnson to the market for LBJ biographies, saying that while Robert Carro’s Master of the Senate would be unaffected, lesser biographies probably would be. But what he forgets is that Master of the Senate has also decreased the market size of competing works too, including biographies that Carro relied upon! Judge Chhabria tries to create a new theory of market dilution, seemingly borrowing from trademark law, but this concept does not seem transferrable to me.
The main way that an AI model would compete with lesser-known, up-and-coming creatives would seem to me to be for works that do not yet exist. For example, the header article for this image was made by an OpenAI image generation model. It does not exist anywhere, but in the past perhaps I would have paid a smalltime artist to create it for me. There is a real market dislocation here with interesting social implications, but copyright does not protect works that do not yet exist, even though Judge Chhabria might want a new harm to be created for “indirect” substitution, contra Warhol’s clear directive that only direct substitution is protected.5
Both cases are quite different than the recent draft guidance from the United States Copyright Office. The prepublication Report was not exactly subtle in suggesting its authors’ view that training is not fair use, though it did steer clear of advocacy, saying that “it is for the courts to weigh the statutory factors.” The Report it did not come to any firm conclusions on any of the four factors, but it did make it clear that, in USCO’s view, any fair use analysis will prioritize factors one and four, and went out of its way to highlight public commentary that put forth arguments that AI training was not entirely transformative and that there would likely be a significant commercial effect on the works.
These cases are instructive when comparing to Thomson Reuters. In that case, Ross created headers very similar to the Westlaw headers for the purpose of creating a model that would form the basis of a competitive service after they approached Thomson Reuters about a license for the headers, which was rejected. This recalls cases like Fox News Network, LLC v. TVEyes, Inc., No. 15-3885 (2d Cir. 2018) and Associated Press v. Meltwater U.S. Holdings, Inc. 931 F. Supp. 2d 537 (S.D.N.Y. March 21, 2013), where if you are trying to use fair use to create a directly competing service—or, in the case of TVEyes, create a service that makes it just a little too easy to recreate a copyrighted work for the purpose of not paying for it. That is clearly not the goal here, where companies treat so-called “regurgitation” as a bug they work to eliminate. Indeed, in Meta Platforms, the authors could not get Meta’s Llama model to produce more than 50 words of their works.6
This is clearly a win for the model providers. These cases clarify that they can train on and digitize data as long as it is lawfully obtained. Although these are district court cases, because it hews so closely to existing precedent, I imagine that it is unlikely to be overturned and that we will see this become one of the seminal cases on the matter. This is exactly the result you would expect if you thought AI was a similar enough technology to other digital technologies to apply cases like Google Books, TVEyes, and Meltwater in an analogous fashion. The reason to believe otherwise would be if you believed AI was fundamentally different than digitization or search in a meaningful way. To me, I have been calling this the TVEyes divide in discussions with my colleagues precisely because this is where I think the dividing line will be: if the service is meant to create a generally useful tool or a clearly competing service. And given that almost all of the AI fair use cases have been dismissed on the substance, these cases might be all we get.
I believe that this is also a win for content owners, however. Just like with the early Internet, the various content industries like music publishing and movie studios began by fighting technological change. This worked poorly for them. Once they started to work with technology companies instead of opposing the very idea of technology, they began to make even more money than before. As an example, music streaming is more profitable than buying records ever was. While it is true that these cases put an implicit cap on the market for licensing content—specifically, at the cost of a single copy of the work—at the same time it makes the market reasonably sized and tractable, thus clarifying it and, I suspect, therefore catalyzing it.
What’s the upshot here? Copyright was always a somewhat limited option for artists to get paid by AI model providers given the strong fair use case. These cases put the spotlight on other IP rights, like trademark and likeness. They also encourage rightsholders to negotiation and practical solutions, like monetizing through AI-generated merchandise and derivative works. I have long been beating the drum, including on Nonobvious, that this was the path forward. I fear that this holding will encourage archiving at the expense of the open web, and that licensing terms for creative works may come with more limitations (unenforceable as they may be). But they will have also started to set the metes and bounds of a real commercial solution to this problem.
Just in time manufacturing was great for rightsholders. Perhaps just in time content will prove to be just as amazing.
Or perhaps I should say, in the AIr?
Sometimes, other ideas can unofficially slip in. In the torrenting cases, for example, intent to pirate seems to have been an unofficial factor. The Napster court went out of its way, for example, to describe the apparent disregard the company had for the law.
For what it is worth, I come down on the side of Judge Alsup. Infringement is infringement. If you want to digitize a copy, you should have to go through the effort of doing so yourself.
As a reader, I found it interesting how fair use cases get passed around regardless of which circuit it is in. The N.D.Cal cases cite Second Circuit law, the D.Del. cases cite Ninth Circuit law, and so on. It truly makes this feel like courts are in a race to define AI precedents for the whole country, not just their circuits, given how evidently persuasive fair use precedents seem to be.
Interestingly, derivative works do not appear at all in either case, even in the discussion. To me, this is probably the most compelling risk on the matter of fair use, but then one needs to dive into the issue of whether the LLM was a tool used by an infringer, in which case there is no liability, or if the AI was an author. The reason derivative works are not covered is partially for this reason, and partially because the question is whether the LLM itself is fair use, not whether its output is fair use, but I was still expecting at least some discussion.