The LLMs Go to Court

Last week two San Francisco judges issued important rulings about the use of copyrighted works to train large language models. Each case involved authors suing a company that offers a large language model (LMM) as service—Anthropic in one case and Meta in the other. In both cases, the authors alleged that the AI companies had violated their copyrights by using their books without permission to train the LLMs. In both cases, the judges sided mostly with the AI companies. But there were significant differences in the judges’ logic that points to two very different outcomes for the future of LLMs.

The underlying facts are largely the same. Both Anthropic and Meta have built LLMs, software models that use massive databases of information to generate new text in response to questions or prompts from users. Creating an LLM requires “training” the model by feeding it huge amounts of text its responses should emulate (in both style and substance). Anthropic and Meta used copyrighted books, among other sources, to train their LLMs. The authors of some of those books claimed that training the models on their works infringed their copyrights. Anthropic and Meta both claimed that their use of the books fell under the fair use exception to copyright protection.

The law says that “fair use” of a copyrighted work doesn’t infringe a copyright. It gives several examples of fair use, like criticism, news reporting, teaching, and research. In other cases, the law directs courts to include four factors when analyzing whether any particular use of a copyrighted work is fair:

1. the purpose and character of the use, including whether it’s of a commercial nature (versus nonprofit or educational);
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 on the potential market for or value of the copyrighted work.¹

The full text of both opinions is available. Here are the key takeaways.

Training an LLM is transformative

The first fair-use factor is the purpose and character of the use. While the statute specifically calls out whether the use is commercial, judges also focus on whether use of the copyrighted material is transformative: whether the copier is just replacing the original with something similar or complementing it with something new and creative. Both judges agreed that this factor favored the AI companies because training an LLM that can generate new text is highly transformative. Judge William Alsup wrote in Anthropic:

The technology at issue was among the most transformative many of us will see in our lifetimes.

That’s an unusual way to interpret the word “transformative” in this context. “Transformative” in copyright law focuses on whether the new work serves a different purpose than the original (common examples are parody, scholarship, and criticism). But here Alsup suggests that the impact on society could be part of the calculus.

And here’s Judge Vince Chhabria in Meta:

There is no serious question that Meta’s use of the plaintiffs’ books had a “further purpose” and “different character” than the books—that it was highly transformative.

Despite the fact that both judges found that the companies’ use was commercial in nature, the training of an LLM is so transformative (and so different from the ordinary purpose of writing or selling a book) that the first copyright factor nonetheless favored the companies.

The Nature of the Work Favors the Authors, but it Probably Doesn’t Matter

By contrast, both judges found that the second fair-use factor, the nature of the work, favored the authors. Both opinions mention that the AI companies valued books because they contain high-quality writing of the type they want their LLMs to emulate and creative ideas that could expand the models’ horizons. But both judges also indicated that this factor didn’t mean much in the overall analysis, Alsup because he found that the other three factors all favored the companies, and Chhabria because the second factor isn’t that important. Chhabria wrote:

The second factor, however, has rarely played a significant role in the determination of a fair use dispute.… So the fact that the second factor favors the plaintiffs doesn’t mean much for the analysis as a whole.

Even though finding high-quality training material is immensely important for creating a good LLM, the judges simply didn’t accord this factor much weight.

Necessity Favors the AI Companies

The third factor evaluates the amount and substantiality of the work that was copied. It’s undisputed that the AI companies copied the authors’ entire works, but the judges didn’t find that fact particularly convincing. Indeed, the judges both held that this factor favored the AI companies for two reasons.

First, the judges found that LLMs need large amounts of text for training, and that the high-quality nature of the authors’ works made their books valuable content for training LLMs. Therefore, Alsup and Chhabria agreed that copying the books in their entirety was necessary for the particular use of training an LLM. Chhabria noted, “Everyone agrees that LLMs work better if they are trained on more high-quality material.” And Alsup concurred, “Because using so many works was reasonably necessary, using any one work for actually training LLMs was about as reasonable as the next.”

Outputs

In evaluating this third factor, the judges also agreed that limitations on the outputs of the LLMs was important. Even though Anthropic and Meta’s LLMs were trained on the authors’ entire books, both companies used filters or similar limitations on outputs to ensure that their consumer-facing products wouldn’t output (at least verbatim) lengthy sections of the authors’ books. And in neither case did the authors establish a record showing that the LLMs had, or could, create copies of their books. As Alsup observed:

When each LLM was put into a public-facing version…, it was complemented by other software that filtered user inputs to the LLM and filtered outputs from the LLM back to the user. As a result, Authors do not allege that any infringing copy of their works was or would ever be provided to users by the [user-facing] service.

And Chhabria:

Given that Meta’s LLMs won’t output any meaningful amount of the plaintiffs’ books, it’s not clear how or why Meta’s copying would be less likely to lead to the creation of direct substitutes for the books if Meta had copied less of them.

It’s worth pausing here to think about the implications of this argument in the broader context of copyright suits against AI companies. The judges say that what matters for the third fair-use factor is whether the user-facing products actually output works similar to the authors’ books. They seem to view the LLMs as black boxes—the fact that they’re unlikely to spit out copies of the authors’ writing means that the “substantiality” factor favors the AI companies, even if the LLMs are trained on, or even store, the entirety of the authors’ books.

Training a model that doesn’t regurgitate the text it’s trained on is a crucial aspect of training a good LLM, but one ramification here (and Alsup calls it out directly) is that filters to prevent an LLM from outputting copied text are more important than the LLM itself. That is, even if a model is capable of producing output that is substantially identical to its training materials, filters that check for copyrighted output could nonetheless help an AI product avoid liability for infringement.

Recently a group of scholars from Stanford, Cornell, and West Virginia University published a paper evaluating whether certain LLMs memorize books they’re trained on (here’s a good article summarizing the study, and the full text). The study finds that some LLMs, including Meta’s Llama 3 at issue in that case, do indeed memorize substantial portions of certain books (almost all of Harry Potter and the Sorcerer’s Stone). The study was published too recently to be relied upon in these cases, but its implications are important nonetheless.

Both judges brushed aside the possibility that the LLMs themselves might be infringing works, citing cases holding that what matters for this analysis isn’t the amount of the work that’s copied, but the amount that’s eventually made available to the public. But the study authors point out that some LLMs, including Meta’s Llama, are actually distributed to users. That is, anyone can download the Llama model and run copies of it themselves. As a result, each download of the model could be viewed as an infringement:

If we say conservatively that the model has been downloaded 1 million times since its release, then those 1 million downloads could be seen as 1 million, potentially infringing distributions of reproductions of Harry Potter and the Sorcerer’s Stone.

This issue wasn’t raised in either case, so the judges didn’t explore it. But it could well be an important factor in future cases, especially for suits against companies that distribute open-weights models, like Meta and Deep Seek. It also casts doubts on the way AI companies present their models. OpenAI, for example, says that “memorization is a rare failure of the learning process.”

Disagreement on the Market Effect

The last fair-use factor—the effect on the market or value of the copyrighted work—is where the judges diverged. AI companies sometimes argue that the process of training a model is like teaching a human being. If someone reads a bunch of books, is influenced by them, and then writes an original novel, we wouldn’t even consider the possibility that it’s copyright infringement. That people might buy this hypothetical writer’s books rather than the books that influenced him or her doesn’t even enter the picture. An LLM learning from copyrighted books that produces output not substantially similar to the original books is essentially the same thing, the argument goes.

Alsup was convinced by this argument, writing:

[The authors’] complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works.

Chhabria took a very different tack. He pointed out that there are at least three ways to argue that LLM outputs displace the market for copyrighted materials the models are trained on:

1. The outputs are substantially similar to the original works, and people will use the LLMs to generate copies of the original works
2. The authors missed out on the opportunity to license the books to the AI companies
3. The advent of LLMs will result in an explosion of new books that people will buy, or get for free, instead of buying the authors'

Chhabria dismisses the first two arguments (the first isn’t supported by the record, and the second is circular and has been largely rejected in other copyright cases). Notably, in Meta, there was no evidence on the record that Llama had memorized, or would regurgitate, any more than about 50 words from the authors’ books. Even the authors’ expert had to admit that Llama wasn’t able to reproduce any “significant percentage” of the works.

But even if the authors had benefited from analysis like the Stanford/Cornell/WVU study they probably wouldn’t have had a stronger case. The amount Llama memorized varied greatly from book to book, though it almost entirely memorized Harry Potter and 1984. So maybe J.K. Rowling and George Orwell would have better cases. Even so, the study authors had to do a lot of work and prompting which they admit was very expensive (in terms of the compute required), so this almost certainly wouldn’t be a time- or cost-effective way to get a copy of a book. The method the study used to test the models’ memorization also depended on knowing what was in the book in the first place. The memorization study is more about what’s inside the models than what they’re likely to outupt in a real-world situation.

But Chhabria finds the third market displacement argument compelling. Chhabria argues an LLM can produce so much more content than a human writer that, regardless of the analogy the AI companies push, training an LLM is a process that’s different in kind from teaching a human being to write:

This case… involves a technology that can generate literally millions of secondary works, with a miniscule fraction of the time and creativity used to create the original works it was trained on. No other use—whether it’s the creation of a single secondary work or the creation of other digital tools—has anything near the potential to flood the market with competing works the way that LLM training does.

There’s a fundamental disagreement here. Alsup thinks the risk of an explosion of AI works that displace human-created books isn’t something we should consider when evaluating copyright claims: “This is not the kind of competitive or creative displacement that concerns the Copyright Act.” Chhabria strongly disagrees, writing that courts have previously recognized the notion of indirect substitution and that LLMs are so different from anything that’s existed before that the law should recognize their potential to upend the market for creative works. He even feels that this risk will move the needle dispositively in the authors’ favor:

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.

Nonetheless, Chhabria ruled for Meta, but only because the authors’ lawyers didn’t really raise this argument and didn’t establish a sufficient record to support it. But Chhabria was clearly disappointed in that outcome, writing that the conclusion of the case “may be in significant tension with reality, but it’s dictated by the choice the plaintiffs made to put forward two flawed theories of market harm.”

Though it wasn’t considered in these cases, the LLM memorization study could also come into play here. Prevalence of memorization by an LLM suggests that LLMs learn very differently than human beings do. (We already sense this intuitively, I think. However human beings learn language and writing, it doesn’t seem like we do so by compiling vast mathematical and statistical models of token association.) I’m no expert on learning or memory, but human beings clearly aren’t capable of memorizing the way computers generally, or LLMs specifically, are. It’s hard to say if or how this distinction should affect LLM copyright cases, but it’s an intriguing difference that wasn’t raised in these cases.

These divergent opinions make it difficult to assess what these outcomes mean for future cases. Authors who find a judge that aligns with Chhabria seem likely to find a sympathetic ear; those who end up before judges that follow Alsup’s logic won’t. The bottom line is that there’s a lot more law to be made on this issue, and plenty of opportunity in other, similar, cases that are pending.

Stealing Books Might be Risky

In both cases, it’s undisputed that the AI companies downloaded pirated copies of the authors’ books. Anthropic and Meta obtained massive databases of stolen books, including the authors’. Both judges analyzed this issue separately from the other copyright infringement claims: Alsup said that the claims regarding Anthropic’s downloading of pirated copies of books will go to trial; Chhabria scheduled a conference to discuss how to proceed on the authors’ similar claim in the Meta case. But both judges found these claims easier to prosecute; claims of pirating are much more familiar. Undoubtedly, AI companies that can build their models with materials obtained legally will have an easier time in court. But whether the pirating amounts to a speeding ticket or something more substantial remains to be seen.