IP Law Daily, COPYRIGHT—N.D. Cal.: Generative AI squeaks out another fair use victory, but will take no comfort in the opinion, (Jun 26, 2025)

By Matthew Hersh, J.D.

A San Francisco judge casts a skeptical eye on the fair use argument—and charts out a different path from his colleague on the bench.

A generative AI model developed by Meta made a fair use of books by a group of fiction and non-fiction authors because its use of the works was highly transformative and because the authors came forward with no evidence that the technology posed a threat to the future market for their work, a federal court in San Francisco has held. But the court, while ruling for Meta on the facts before it, made clear that it would have decided the case differently on a better-developed record—and its opinion may well revitalized both the fourth fair use factor as well as a controversial doctrine of “indirect substitution” that to now has remained only the stuff of law reviews and legal treatises (Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. June 25, 2025)).

The lawsuit arose when 13 authors sued Meta for downloading their books from online “shadow libraries” and using the books to train Meta’s generative AI models. The plaintiffs, whose ranks include non-fiction writer Sarah Silverman, Pulitzer Prize-winning novelists Junot Díaz and Andrew Sean Greer, and many other well-known authors, took aim specifically at Meta’s downloadable large language model known as Llama. The authors filed the lawsuit seeking to represent a class of all owners of copyrighted works used as training data for Llama. They brought claims for direct copyright infringement, vicarious copyright infringement, and removal of copyright management information in violation of the Digital Millennium Copyright Act, along with several common law claims.

Meta moved for summary judgment on its defense of fair use, leading to this opinion.

Transformative use. The court found that Meta made a fair use of the authors’ works. The court’s sweeping opinion addressed all four factors—purpose and nature of the use, nature of the work, amount and proportionality of the taking and market effect. But the court’s analysis turned almost exclusively on the first and fourth factors—particular transformative use and market effect.

The first factor, the court found, favored Meta. “There is no serious question,” the court said, “that Meta’s use of the plaintiffs’ books had a further purpose and different character than the books—that it was highly transformative.” The purpose of Meta’s copying, the court observed, was to train its LLMs, which are “innovative tools that can be used to generate diverse text and perform a wide range of functions.” Thus, the court noted, users can ask Llama to edit an email they have written, translate an excerpt from or into a foreign language, write a skit based on a hypothetical scenario, or do any number of other tasks. The purpose of the plaintiffs’ books, by contrast, the court explained, is “to be read for entertainment or education.”

In finding that Meta made a transformative use of the books, the court rejected an argument—made by law professors writing to support the authors—that Meta’s use had the same purpose and character as the books “because an LLM training on a book is akin to a human reading one.” This argument would fail, the court noted, because “an LLM’s consumption of a book is different than a person’s.” An LLM ingests text, the court noted, to learn statistical patterns of how words are used together in different contexts—making billions of calculations based on the date in its training set. “This is not how a human reads a book,” the court said.

Nor did it matter, according to the court, that Llama had “no critical bearing” on their books, the way criticism or parody would. That was true, the court noted—but also irrelevant. Importantly, the court found, critique or commentary on the original are “not the only uses that will furnish a justification ultimately qualifying as fair use.” To the contrary, a use that enables “the furnishing of valuable information on any subject of public interest” or renders “a valuable service to the public” might be justified, the court noted, especially where that benefit is “provided without allowing public access to the copy.”

Nature of the work and amount and proportionality of the use. The court issued a split decision on these factors, although neither weighed much in the analysis. The nature of the authors’ works was a factor tipping in favor of the authors, the court found, because their books—mostly novels, memoirs, and plays—were “highly expressive works of the type that the copyright laws value and seek to protect.” Moreover, the court noted, Meta used these books not merely to gain access to their “functional” elements, but to capitalize on their creative expression—indeed, the evidence showed that Meta highly valued well-written long-form books as training materials. Nonetheless, the court noted, the second factor rarely plays a significant role in the determination of fair use—especially where, as here, the works in question have already been published.

The third factor, by contrast, favored Meta—although it, too, would not play much of a factor in the analysis. To be sure, the court observed, Meta copied the authors’ books in their entirety. But that copying, the court found, was reasonable given its relationship to Meta’s transformative purpose. “Everyone agrees that LLMs work better if they are trained on more high-quality material,” the court noted. “So feeding a whole book to an LLM does more to train it than would feeding it only half of that book.” With this in mind, the court concluded, it was “reasonably necessary for Meta to make use of the entirety of the works.”

Market effect. The fourth factor, too, weighed in Meta’s favor, the court found. The court began by emphasizing the importance of the fourth factor. “Meta is wrong to suggest,” the court noted, “that because the first factor strongly favors it, the inquiry should basically end there.” To the contrary, given the fourth factor’s importance, the court noted, “it’s easy to imagine a situation in which a secondary use is highly transformative but the secondary user nonetheless loses on fair use because allowing people to engage in that kind of use would have too great an effect on the market for the original work.” But by the same token, the court noted, “in a case where the first factor cuts strongly in favor of the defendant, generally the plaintiff’s only chance to defeat fair use will be to win decisively on factor four.”

Here, the court concluded, the authors failed to muster that decisive win on factor four. The authors rested primarily on two arguments for why Llama posed a threat to the market for their works—both of which the court handily rejected. First, the authors argued that there was a risk that Llama might “regurgitate their works (or outputs that are substantially similar),” thereby allowing users to access those works or substitutes for them for free.” But Llama simply did not do that, the court found—indeed, neither party’s expert opined that Llama was able to regurgitate more than 50 words from any of the plaintiffs’ books. The authors also argued that Meta’s unauthorized use of their books for LLM training harmed the market for licensing their books for that purpose. But that argument, too, would fail. “Whether such a market exists or is likely to develop is irrelevant,” the court reasoned, “because this market is not one that the plaintiffs are legally entitled to monopolize.” The argument, indeed, was “circular,” the court noted—in every fair use case, the “plaintiff suffers a loss of a potential market if that potential market is defined as the theoretical market for licensing the use at issue in the case.” The argument would fail here.

The more promising argument for market effect, the court noted, was the market dilution effect—but here, the court found, the authors simply failed to make a factual record to support their case. An LLM could eventually be used, the court noted, “to generate massive amounts of text in significantly less time than it would take to write that text, and using a fraction of the creativity.” Given that possibility, the court noted, “people could thus use LLMs to create books and then sell them, competing with books written by human authors for sales and attention.” But the problem with this argument, the court noted, was that the plaintiffs simply never developed it on the record. Indeed, the court noted, “the plaintiffs never so much as mentioned it in their complaint,” nor in their summary judgment motion. On the record in this case, the court concluded, Meta had “defeated the plaintiffs’ half-hearted argument” that its copying threatens significant market harm. “That conclusion may be in significant tension with reality,” the court concluded, “but it’s dictated by the choice the plaintiffs made to put forward two flawed theories of market harm while failing to present meaningful evidence on the effect of training LLMs like Llama with their books on the market for those books”

Distribution. The court’s order left unresolved one aspect of the authors’ copyright infringement claim—that having to do with Meta’s alleged distribution of their work through peer-to-peer networks. The authors alleged that Meta downloaded their works by using BitTorrent, a peer-to-peer network that, unless calibrated differently, by default also re-shares downloaded works with other users. But “the record on Meta’s alleged distribution is incomplete,” the court noted, “making summary judgment on that issue improper at this point in the case.”

DMCA claims. Finally, albeit in brief, the court also noted that it would deny the authors’ claims under the Digital Millennium Copyright Act, for reasons that it would articulate in a later opinion.

Commentary. The court’s opinion is striking in many ways. Most importantly, Judge Chhabria took an entirely different read of fair use—especially the market value prong of the analysis—than his colleague on the San Francisco bench did just two days ago in the Anthropic opinion. In that case, in granting summary judgment to the developers of AI chatbox Claude, Judge Alsup was largely dismissive of the harm that generative AI might posed to the market for the works it gets trained on. Such harm would be no different, Judge Alsup found, “than the harm caused by using the works for “training schoolchildren to write well,” which could “result in an explosion of competing works.” In language perhaps unusual for its directness, Judge Chhabria called this an “inapt analogy.” When it comes to market effects, Judge Chhabria emphasized, “using books to teach children to write is not remotely like using books to create a product that a single individual could employ to generate countless competing works with a miniscule fraction of the time and creativity it would otherwise take.” The analogy was “not a basis for blowing off the most important factor in the fair use analysis,” he concluded.

Judge Chhabria’s opinion was also striking because it ventured into what the Copyright Office labeled just two months ago the “uncharted territory” of the “market dilution” effect—or “indirect substitution,” as it is also described. Traditional fourth factor analysis has focus only on the harm of market substitution—that is, the concern over whether a market for the original work is supplanted. But in an unusual “pre-publication” release of a controversial Copyright Office report on artificial intelligence—rushed out just before President Trump’s firing of the head of that office—the Office argued that courts should “consider the harms caused where a generative AI model’s outputs, even if not substantially similar to a specific copyrighted work, compete in the market for that type of work.” Judge Chhabria, while (understandably) not specifically referencing the report, nonetheless wholeheartedly endorsed its conclusion. “Indirect substitution is still substitution,” Judge Chhabria insisted. “If someone bought a romance novel written by an LLM instead of a romance novel written by a human author, the LLM-generated novel is substituting for the human-written one.” That kind of indirect substation might not be enough to move the needle in a run-of-the-mine copyright infringement case, Judge Chhabria noted—but artificial intelligence was different. “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,” the court noted. “And so the concept of market dilution becomes highly relevant.”

Finally, Judge Chhabria’s opinion was striking because, while it granted a victory to Meta on the particular facts, it made very clear that on a more developed record, the developers of generative LLMs would likely lose. “As should now be clear,” the court noted, “this ruling does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful. It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.” The upshot, the court noted, is that “in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission—[w]hich means that the companies, to avoid liability for copyright infringement, will generally need to pay copyright holders for the right to use their materials.”

The Case is No. 3:23-cv-03417-VC.

Judge: Chhabria, V.

Attorneys: Joseph R. Saveri (Joseph Saveri Law Firm, LLP) for Richard Kadrey. Bobby A. Ghajar (Cooley LLP) for Meta Platforms, Inc.

Companies: Meta Platforms, Inc.

MainStory: TopStory AINews Copyright TechnologyInternet CaliforniaNews GCNNews