Key Rulings on GenAI Training and Copyright Fair Use Practical Law The Journal

Reuters

a day ago

For a regularly updated case tracker covering intellectual property and privacy-related lawsuits concerning GenAI (including more decisions addressing fair use), see Generative AI: Federal Litigation Tracker on Practical Law.
Bartz v. Anthropic PBC: N.D. Cal.
On June 23, 2025, the US District Court for the Northern District of California held in Bartz v. Anthropic PBC that defendant Anthropic PBC's use of copyrighted books to train its GenAI tool was a fair use and granted summary judgment to Anthropic on this issue. The court also held that Anthropic's digital conversion of purchased print books to build its digital library was fair use but that its downloading of pirated copies for this purpose was not. (2025 WL 1741691 (N.D. Cal. June 23, 2025).)
Anthropic PBC developed the GenAI tool Claude, which generates text responses based on prompts from users. In part to train the large language models (LLMs) underlying Claude, Anthropic assembled a central library of digitized books, including copies purchased in print form and then scanned into digital format, as well as copies downloaded from pirate websites.
Authors Andrea Bartz, Charles Graeber, Kirk Wallace Johnson, and their affiliated corporate entities (collectively referred to as the 'Authors') filed a putative class action lawsuit against Anthropic in August 2024, alleging copyright infringement for using copies of their books to build its digital library and train the LLMs. Anthropic moved for summary judgment on the issue of fair use. The district court analyzed the fair use factors under Section 107 of the Copyright Act to determine whether Anthropic's uses of the Authors' copyrighted works constituted fair use, separately evaluating the different uses at issue.
Weighing the factors, the Bartz court concluded that Anthropic's use of the Authors' books to train the LLMs was fair use. Specifically, the court found that:
The first factor (purpose and character of the use) strongly favored fair use because using the works to train the LLMs to generate new text outputs was 'quintessentially transformative.' Key to this finding was that Claude includes software to ensure that it does not output infringing content (and the Authors did not allege that any output was infringing). The court acknowledged that its analysis may change if the outputs were infringing.
The second factor (nature of the copyrighted work) weighed against fair use because the court accepted, at the summary judgment stage, that:
the Authors' published works contained expressive elements; and
the works were selected for their expressive qualities.
The third factor (amount and substantiality of the portion used) favored fair use because, although Anthropic copied the Authors' entire works, this was reasonably necessary given the extensive data needed to train the LLMs. The court also stated that what matters is the amount and substantiality of what is made accessible to the public, again noting that there was no allegation or evidence that consumer-facing outputs were infringing.
The fourth factor (effect on the potential market) favored fair use because:
The district court also considered whether Anthropic's copying of the Authors' works to build its central digital library was fair use. The court separately considered works that were:
Lawfully purchased in print format and converted to digital format, after which the print versions were destroyed and the digital versions were not redistributed.
Copied from pirate websites without authorization by or compensation to the Authors.
For the purchased print copies, the district court held that their conversion for a digital library was a fair use. The court found:
The first factor favored fair use because:
converting the works from physical to digital format for storage and searchability was a transformative use; and
Anthropic did not create additional copies or redistribute the digital versions.
The second factor weighed against fair use based on the presumptively (at the summary judgment stage) expressive nature of the works.
The third factor favored fair use because copying the entire work was necessary for the purpose of digital conversion and storage.
The fourth factor was neutral, as the format change may have displaced some digital purchases, but this did not relate to a market the Copyright Act entitles the Authors to exploit.
However, for the pirated library copies downloaded without authorization, the district court found no fair use justification and denied summary judgment to Anthropic. Anthropic copied these pirated works, as a substitute to purchasing them, to build a digital library available for any number of prospective uses (and maintained copies in the library even after deciding they would not be used to train the LLMs). The court held this use is not transformative. The court further recognized that the pirated copies directly displaced demand for purchased copies on a one-to-one basis and that condoning such piracy as fair use would destroy the publishing market. The court rejected Anthropic's arguments that the eventual transformative use of some copies for training the LLMs excused the initial piracy.
Kadrey v. Meta Platforms, Inc.: N.D. Cal.
On June 25, 2025, the US District Court for the Northern District of California granted summary judgment in Kadrey v. Meta Platforms, Inc. to Meta Platforms, Inc., finding that Meta's use of the plaintiffs' copyrighted books to train its GenAI tool was transformative and fair use. However, the court stated its belief that the fair use defense is likely to be unsuccessful in other, similar cases where the copyright owners adequately show the dilutive harm that GenAI has on the general market for these works. (2025 WL 1752484 (N.D. Cal. June 25, 2025).)
The plaintiffs, thirteen authors, filed a lawsuit against Meta alleging direct copyright infringement, among other claims, based on Meta's use of unauthorized downloads of their books (from online shadow libraries) to train the LLMs underlying Llama, Meta's text-generating GenAI tool. After discovery, the parties cross-moved for summary judgment on whether Meta's use of the books was fair use. The district court analyzed the fair use factors under Section 107 of the Copyright Act (17 U.S.C. § 107).
In support of its finding that the first fair use factor (purpose and character of the use) favored fair use, the district court:
Held that Meta's use was highly transformative because it used the plaintiffs' books only to train LLMs, while the purpose of the books is to entertain and educate readers. The Kadrey court noted that transformative use does not insulate a defendant from infringement liability or even determine the first fair use factor. It is one aspect of the fair use analysis, and there are circumstances where market harm (the fourth factor) can be grounds to reject the defense for a transformative use.
Rejected the plaintiffs' law professor amici argument that the purpose and character of the parties' uses were similar because Meta's use of a book to train the LLMs was like a professor's use of a book to train a student. The district court noted that:
an LLM ingests text only to learn statistical patterns, not to interpret and understand its meaning as a student does; and
Meta's use was not analogous to giving a book to one person, but rather it was to create a tool that everyone can use to exponentially multiply creative expression in a way that teaching a person does not.
Rejected the plaintiffs' argument that Meta's use was not transformative because Llama's output mimics and effectively repackages the plaintiffs' works. The evidence showed that Meta programmed Llama to be unable to regurgitate training content, and the plaintiffs' experts were unable to prompt the tool to generate more than 50 words from any of the plaintiffs' books.
Recognized that Meta's commercial use (and expectation of 460 billion to 1.4 trillion dollars in revenue over the next ten years) tends to weigh against fair use, but this did not tilt the first fair use factor in the plaintiffs' favor.
Recognized that Meta's unauthorized downloading of the books from shadow libraries without compensation to the plaintiffs may indicate bad faith, but questioned the relevance of bad faith and found it did not sway the first factor in this case. The court noted that Meta's practice might be more relevant to the character of the use if the plaintiffs showed the practice benefited the shadow libraries and furthered their unlawful activities.
The court held that the second factor (nature of the copyrighted work) weighed against fair use because the plaintiffs' books, consisting mostly of highly expressive works, are entitled to strong copyright protection. However, the court noted that this factor rarely plays a significant role in the fair use analysis.
The district court acknowledged that the third factor (amount and substantiality of the portion used) was not particularly relevant in this case. However, it concluded that the factor favored a fair use finding because copying the entirety of the books was reasonable given Meta's transformative purpose, as LLMs perform better when trained on complete, high-quality data.
The district court started its review of the fourth fair use factor (effect on the potential market for the copyrighted work) by acknowledging it to be the most important factor in the fair use analysis. It explained that the relevant question is whether the defendant's use will function as a market substitute for the plaintiffs' works. The court rejected the plaintiffs' arguments that:
Meta's unauthorized use of the plaintiffs' books affects the market for licensing the works for the purpose of training its LLMs. The district court held that this is not a harm that the Copyright Act seeks to prevent. Otherwise, every copyright infringement plaintiff could argue that it has been deprived of the right to license the use at issue in the case.
Llama is capable of reproducing portions of their books, therefore harming the market for the plaintiffs' works. However, the evidence showed that even adversarial prompting designed to make Llama regurgitate the plaintiffs' works yielded only 50-word snippets from the books, which could not have a meaningful effect on the market for the plaintiffs' books.
Although the plaintiffs focused only on these two alleged harms, the district court analyzed a third form of harm, that is, GenAI's ability to rapidly generate countless works that compete with and reduce demand for the plaintiffs' works, even if the AI-generated works are non-infringing. The court referred to this form of harm as market dilution (or indirect substitution), which it noted is still market substitution, could reduce the incentive for authors to create, and is the specific harm that copyright aims to prevent. The court stated that market dilution harm is far greater (and therefore more relevant) in the case of GenAI than in other cases involving individual secondary works or digital tools, such as Google Books, because GenAI can quickly flood the market with millions of competing works.
The court stated that it 'seems likely' that market dilution will often cause the fourth fair use factor to decisively favor plaintiffs in similar cases. However, in this case, because Meta introduced evidence that its use of the plaintiffs' works did not cause market harm and the plaintiffs failed to demonstrate the contrary, the district court (seemingly reluctantly) found that the fourth factor weighed in favor of fair use.
In granting summary judgment to Meta on its fair use defense, the district court stated that its ruling should not indicate that Meta's use of copyrighted content to train its LLMs was lawful, but only that the plaintiffs did not show the market dilution that GenAI causes. The court further surmised that, in many circumstances, the unauthorized use of copyright-protected works to train GenAI models will be infringing and developers will therefore need to pay copyright owners for the right to use their materials for this purpose.