Two judges in the Northern District of California recently issued groundbreaking summary judgment rulings regarding whether an artificial intelligence company’s scraping and ingestion of copyrighted works to train its LLMs[1] qualified as fair use. Both decisions carry potentially seismic importance for AI companies and intellectual property litigators.
Overview of the Anthropic and Meta Rulings
The first ruling arose in a proposed class action of book authors who alleged that Anthropic violated their copyrights to build its LLM, Claude. Bartz, et. al. v. Anthropic PBC, Case No. C-24-05417 (N.D. Cal.). The complaint alleges that Anthropic built its library of training materials in part by purchasing copyrighted works in hard copy form, then scanning their pages and storing the digital versions. Anthropic also allegedly downloaded copyrighted works directly from pirate websites and trained Claude on this library of collected works.
Anthropic moved for summary judgment on its fair use defense. To decide whether use of a copyrighted work is “fair use,” courts look at four factors:
- The purpose and character of the use, including whether the use is for commercial or nonprofit educational purposes;
- The nature of the copyrighted work;
- The amount and substantiality of the portion of the work that is used;
- The effect of the use upon the potential market for or value of the copyrighted work.
In his June 23, 2025 ruling granting Anthopic’s motion and finding fair use (ruling available here), Judge Alsup first separated Anthropic’s “uses” of training the LLM versus building a central library of content. Assessing the first prong of the fair use test, he ruled that the use of the copyrighted works to train Claude “was exceedingly transformative and was a fair use.” As to the other use—digitizing the works to create a central library—Judge Alsup again found fair use because this involved simply putting the works in a more convenient format “without adding new copies, creating new works, or redistributing existing copies.” In a win for the authors, however, Judge Alsup criticized Anthropic for downloading the copyrighted works from pirate websites and permanently storing them in Anthropic’s central library, ruling that this was not fair use and was “inherently, irredeemably infringing.”
Just two days later, a different Northern District judge (Judge Chhabria) issued his own summary judgment ruling in a similar case brought against Meta by famous authors who alleged that Meta had downloaded their books from online “shadow libraries” to train the company’s Llama AI tool. See Kadrey, et al. v. Meta Platforms, Inc. Case No. 23-cv-03417 (N.D. Cal.) (order available here). In this case, both sides moved for summary judgment on fair use, so Judge Chhabria grappled with the same four factors. While Judge Alsup focused more on the first factor, Judge Chhabria emphasized the fourth factor of market impact because an AI could create cheaper expressive works and reduce human beings’ incentive to create. He noted that AI can “generate countless competing works with a minuscule fraction of the time and creativity it would otherwise take.” He also found, though, that “there’s no disputing that” “the companies’ use of the works is transformative.” He reluctantly granted summary judgment of fair use to Meta, writing that the Court “had no choice” but to do so, but cautioning that “the consequences of this ruling are limited.” He explained that the outcome was “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.”
Both these high-profile cases involved book authors suing AI companies for using their copyrighted works to train their LLMs, and in both, the AI companies had attempted to license the copyrighted works before opting for other methods—including downloading copyrighted works from pirate websites. And in both cases, the judges found fair use. But, as discussed below, the rulings differ in important respects in their tone, outcomes, and what they might mean for the future of fair use in the AI context.
Comparing the Courts’ Fair Use Analyses
Factor One: The purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes.
The courts took different approaches to how the “uses” should be analyzed. Judge Alsup separated the uses of (1) training the LLM and (2) building a library of copyrighted texts, rejecting Anthropic’s argument that they were all part and parcel of the same use. Judge Chhabria, on the other hand, considered these same two “uses” as part of a single analysis because downloading the books had to be considered “in light of its ultimate, highly transformative purpose: training Llama.”
Regardless, the judges agreed that using copyrighted works to train AI systems was “transformative,” and it wasn’t even a close call. Judge Alsup wrote that “the ‘purpose and character’ of using works to train LLMs was transformative—spectacularly so.” Judge Chhabria agreed that Meta’s use of the works “had a ‘further purpose’ and ‘different character’ than the books—that it was highly transformative.” Still, Judge Chhabria explained that the transformative purpose was not the full fair use (or even factor one) analysis, and other considerations should be taken into account—like the commercial nature of the copying, which was relevant and should not be brushed aside. Still, neither these factors nor Meta’s use of “shadow libraries” were significant enough to overcome the highly transformative nature of AI at factor one.
Judge Alsup, in analyzing Anthropic’s building of its library as a separate use, drew a distinction between the works Anthropic purchased and those it pirated. He ruled that digitizing purchased hard copy works to improve storage and searchability and storing those digital copies was a fair use, particularly because Anthropic destroyed the original hard copy and did not share the digital copy with others. But on the pirated works, he took a different tack: “[T]he person who copies the textbook from a pirate site has infringed already, full stop,” he stated, noting that the copies were otherwise available for purchase. Where the “piracy was the point,” and where Anthropic downloaded the full-text copies to be maintained forever, Anthropic’s arguments for fair use failed.
The differences in the courts’ two approaches on this seemingly simple question (Are downloading and training a single use, or two?) are emblematic of the challenge that litigants and courts have found in trying to apply AI technology to traditional copyright infringement case law. This question has potentially large implications for how courts address this issue in the future. If building the library is a use standing by itself, the argument for transformative use is much weaker.
Factor Two: The nature of the copyrighted work.
Both judges devoted little attention to factor two, which they agreed weighed against fair use because the authors’ copyrighted works were creative original works of expression that copyright was designed to protect. Judge Alsup explained that factor two is mainly used to assess the other factors, and Judge Chhabria agreed that the second factor “doesn’t mean much for the analysis as a whole.”
Factor Three: The amount and substantiality of the portion used in relation to the copyrighted work as a whole.
The judges came to similar conclusions here regarding training the LLM. Both pointed out that the relevant question is not the amount of copyrighted material used by the copier, but the amount that the copier makes available to the public in the purported transformative use. Here, Judge Alsup ruled that copying entire books was necessary for the transformative use. Anthropic needed billions of words to train an LLM, and AI outputs accessible to the public were not at issue, weighing in favor of fair use. Regarding the authors’ argument that Anthropic did not need to use their specific works to train the model, Judge Alsup found that “[b]ecause using so many works was reasonably necessary, using any one work for actually training LLMs was about as reasonable as the next.” Judge Chhabria similarly concluded that using the whole book to train an AI that needs large volumes of high-quality data is reasonably necessary.
Judge Alsup also examined Anthropic’s alleged use of the copyrighted works in building a library, and again distinguished between purchased and pirated copies. For purchased copies, he found fair use because “[c]opying the entire work was exactly what this purpose required,” “[t]here was no surplus copying,” and “[t]he source copy was destroyed.” However, for pirated copies, he ruled that almost any copying would be too much because Anthropic had no entitlement to them at all, which pointed against fair use.
Factor Four: The effect of the use upon the potential market for or value of the copyrighted work.
The market effect factor reveals the biggest difference between the judges’ views of these disputes. For Judge Alsup, the copies used to train LLMs “did not and will not displace demand for copies of the Authors’ works.” He found that AI and LLM products that could result in an explosion of works competing with original works were no different than schoolchildren learning how to write better by reading books. For the copies used to build a central library, again, he drew the “purchased vs. pirated” distinction. For the purchased hard copies that Anthropic digitized, because it was just a format change and there was no market displacement by buying print rather than electronic versions, this factor was neutral. But for the pirated copies that “plainly displaced demand for Authors’ books—copy for copy,” the factor weighed against fair use.
Judge Chhabria emphasized the importance of factor four much more than Judge Alsup, but acknowledged that because Meta’s use was so transformative, the market impact had to be strongly against fair use to carry the day. He identified three ways that the plaintiff could argue market impact in this context: (1) outputs that are substantially similar; (2) harm to the market for licensing; and (3) outputs that are “similar enough” that they will compete with originals and substitute for them. Here, option (1) failed because Llama did not generate any meaningful portion of the books, and option (2) failed because the argument was circular—the harm from lost licensing fees would always apply to fair use analysis. For the third option that AI would cause “market dilution” by creating works using a fraction of the time and creativity, Judge Chhabria discussed it at length. He explained that although LLMs might not harm famous authors like Agatha Christie, they could prevent the emergence of the next Agatha Christie. This “indirect substitution” for vaguely similar works, while normally negligible, was different here because AI is “a technology that can generate literally millions of secondary works, with a minuscule fraction of the time and creativity used to create the original works it was trained on.” Despite these misgivings, he found that the authors had not put forward enough evidence to create an issue of fact on the “market dilution” theory. But his ruling expressed discomfort with this outcome, noting that his “conclusion may be in significant tension with reality.”
Observations on What (Might) Come Next
We leave readers with five takeaways from this pair of opinions:
- Using copyrighted works to train AI LLMs is “transformative”: Both judges were unequivocal on the transformative purpose of training AI tools with copyrighted works. Judge Alsup emphasized it was “among the most transformative many of us will see in our lifetimes,” and Judge Chhabria stated that “there’s no disputing” its transformative nature. Academics and AI companies have been pushing the courts and the Copyright Office to recognize the transformative nature of AI for a long time.[2] These two rulings from respected and highly analytical judges suggest that courts are likely to agree.
- AI outputs are a different case: AI copyright issues involve both inputs to train LLM models and the LLM’s outputs responding to user prompts. But both these rulings focused on inputs. Judge Alsup noted that if the allegations involved infringing outputs, “this would be a different case” and all but invited plaintiffs to file such a case. Given the difficulty that the plaintiffs in these cases faced in articulating harm to the market, pursuing cases based on output may be the next step for plaintiffs. In fact, the recent complaint filed by Disney and Universal against Midjourney for allegedly generating infringing images of copyrighted characters is an output-focused case.
- Pirating copyrighted works to train an AI is an open question: The two courts addressed pirated works differently. While Judge Alsup found downloading the pirated copies to build a library was an inherently infringing use “full stop,” Judge Chhabria said this fact was “not an automatic win” for plaintiffs because it had to be considered “in light of its ultimate, highly transformative purpose” of training the AI. One could imagine plaintiffs picking up Judge Alsup’s view and pursuing claims based solely on pirated works.
- Factor four might be the crux: The two courts both found that factor four favored fair use, but examined it differently. Judge Alsup found there was no market displacement for the authors’ works (for the training use) because the creation of more works is simply competition, which the Copyright Act does not prohibit. Judge Chhabria focused instead on “market dilution” and, although he could not find for plaintiffs in this case, he stated that “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.” Future plaintiffs will likely sharpen their market dilution theories in an effort to overcome the highly transformative nature of this technology.
- This is not the end of either story: Neither of these rulings disposes of the entire case. The Anthropic case will go forward on the issue of the pirated copies, and there will be further litigation regarding the library copies for uses other than for training LLMs. The Meta case still involves the plaintiffs’ claim that Meta unlawfully distributed their works during the downloading process (a byproduct of “torrenting”). While this was not the core conduct that plaintiffs were targeting when they filed suit, this is a good illustration of the difficulty discussed above—when does one step in the chain of creating an LLM constitute a separate “use,” and when is it just part of the overall process?
These cases are a good reminder that the fair use doctrine is adaptable. Both judges applied the traditional fair use factors to this transformative new technology, but they analyzed similar fact patterns in different ways—highlighting the fluidity of the fair use doctrine. While we cannot know exactly what comes next, it is likely that these thoughtful rulings will form the foundation for future judicial analysis and litigation in the coming years.
Reprinted with permission from the August 31, 2025 edition of the Law.com © 2025 ALM Global Properties, LLC. All rights reserved.
