By: Dana J. Finberg, Partner; C. Quinn Adams, Partner; Madeline Swank, Summer Associate; and Matthew Essien, Summer Associate

It is impossible to ignore the recent developments in artificial intelligence. Technology companies continue to create sophisticated generative AI products like Bard, ChatGPT, Claude, and Gemini, with each coming closer and closer to imitating human intelligence. The backbone of these products is the large language model (LLM). An LLM is a system that can assess data inputs and formulate responses to prompts. Tech companies train these LLMs using terabytes of data to help the LLMs think, reason, and respond like a human. As coal was to the rudimentary machines of the Industrial Revolution, the data fed into LLMs are to the AI Revolution.

But unlike coal, the law affords protections for the “fuel” that tech companies shovel into in their LLMs. Notably, the Copyright Act protects various artists, authors, and others’ original works of creative expression. Authors, artists, and others have increasingly claimed that tech companies’ use of their registered works to train LLMs infringes their federal copyrights. The tech companies respond that this is simply fair use. Courts must consider four factors when assessing a claim of fair use: “(1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; (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 upon the potential market for or value of the copyrighted work.” 17 U.S.C. § 107.

Until June 2025, there was relatively little caselaw addressing this issue.  However, within days of one another two judges in the U.S. District Court for the Northern District of California, Judges Alsup and Chhabria, issued opinions concerning fair use of registered works to train LLMs. Both cases involved authors’ claims that tech companies infringed the authors’ federal copyrights by using their registered works to train LLMs for the companies’ generative AI products. Both companies sought summary judgment on their fair use defense.  Both courts granted the companies’ motions for summary judgment, at least in part.

On the surface, these opinions are consistent with one another. But the devil is in the details.  The differences between Judge Alsup’s and Judge Chhabria’s analysis of transformative use and market dilution reveal the knotty issues that will continue to plague courts as the deal with a technology that is, as Judge Chhabria put it, “both as transformative and as capable of diluting the market for the original works as LLM training is.”

Bartz v. PBC (Judge Alsup)

Several authors brought a putative class actions against Anthropic PBC, for copyright infringement. Anthropic produces the generative AI assistant, Claude. Anthropic built a vast digital library of copyrighted works—some it purchased in hardcopy form and digitized and others it downloaded from pirate sites. Anthropic used this library of works to train the LLMs underlying the Claude product.

The authors argued that Anthropic’s use of the their copyrighted works, including converting the print works into digital form, infringed their federal copyrights. Anthropic contended that its use of the authors’ works was fair use, and moved for partial summary judgment.

Judge Alsup assessed Anthropic’s actions against the four fair use factors. Those actions included: (1) using purchased and pirated copies to train LLMs; (2) using purchased and pirated copies to build a central library; (3) converting purchased hardcopy works into digital form for the central library.

Judge Alsup held that “purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative.” Anthropic did seek to “replicate or supplant” the works used to train its LLMs, but to have those LLMs “turn a hard corner and create something different.” Additionally, Judge Alsup concluded that the extent of Anthropic’s copying was necessary for the transformative use given the “monumental” amount of data needed to train LLMs.  Finally, Judge Alsup concluded that there was no appreciable effect on any market for, or on the value of, the authors’ copyrighted works.

Thus, Judge Alsup concluded Anthropic’s use of purchased copies for the purposes of training the LLMs was fair use. He reached the same conclusion about Anthropic’s use of purchased copies to create a digital library, as well as its conversion of those purchased works from hardcopy to digital form. However, he denied summary judgment concerning Anthropic’s use of pirated copies, including using those copies to create its central library.

In reaching his conclusions, Judge Alsup rejected the authors’ argument that “training LLMs will result in an explosion of works competing with their works — such as by creating alternative summaries of factual events, alternative examples of compelling writing about fictional events, and so on.” He concluded that this 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.”  He concluded that this complaint represented the author’s attempt to use the Copyright Act to protect them from competition, rather than to advance original works of authorship.

Judge Alsup issued a subsequent opinion on July 17, 2025, certifying a nationwide Pirated Books Class, and ordering that notice be provided to putative members of the class.

Kadrey v. Meta Platforms, Inc. (Judge Chhabria)

Several authors sued Meta for obtaining the authors’ copyrighted works from online libraries and using those works to train Meta’s LLMs. The authors and meta filed cross-motions for summary judgment on the issue of fair use.

The authors complained that Meta’s use deprived the authors of their ability to license their works for training AI models and that the Meta’s LLMs could reproduce small portions of their books. Judge Chhabria rejected both arguments because Meta’s LLMs could not produce sufficient material from the authors’ copyrighted works to make a difference and the authors were not entitled to a market for licensing their works for AI training.

Judge Chhabria noted that the authors failed to advance their strongest argument: “that Meta has copied their works to create a product that will likely flood the market with similar works, causing market dilution.”  Because of this, he had “no choice” but to grant summary judgment to Meta on its fair use defense. However, Judge Chhabria took pains to limit his decision, noting his opinion “does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful . . . only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.”

Judge Chhabria also made a point of registering his disagreement with Judge Alsup’s analysis in Bartz. He charged that Judge Alsup with focusing too heavily on “the transformative nature of generative AI while brushing aside concerns about the harm it can inflict on the market for the works it gets trained on.” Judge Chhabria referred to Judge Alsup’s analogy of schoolchildren who write well causing an “explosion of competing works” as “inapt.” He retorted that “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.”

Conclusions

These two decisions reveal the complex legal analysis for applying fair use in AI cases. Some jurists will emphasize the transformative nature of AI products and LLMs. Others will scrutinize the market effects of using copyrighted materials to train LLMs, and will be open to argument that this use stifles the creative expression the Copyright Act is meant to protect. Given the fact-intensive nature of the fair use defense, a hard-and-fast rule on fair use in the context of generative AI is unlikely to emerge. Therefore, practitioners should study particular judges’ prior opinions on fair use (if any) to assess which factors the judge may emphasize in assessing a fair use defense in a case involving generative AI products.