Doctrinal Divergence in AI Copyright: A Fair Use Analysis of Bartz v. Anthropic and Kadrey v. Meta

This blog was written with contributions from Leo Lichtman, Partner at ESCA Legal, and Lea Leisure, Intern. Can a machine “read” a book the same way a person does? And if so, does it owe the author anything for doing so? These questions now sit at the heart of a growing legal debate as copyright claims against AI companies work their way through the courts. In Bartz v. Anthropic and Kadrey v. Meta, both defendants won on summary judgment—but the analytical paths taken by the courts, both sitting in the Northern District of California, diverged sharply. Together, the cases expose deeper tensions in how fair use doctrine applies to generative AI training and suggest that the most consequential questions are still ahead.

Fair Use Framework and the Four-Factor Test

The fair use doctrine explains that certain uses of a copyrighted work are not an infringement. As codified in 17 U.S.C. § 107, it requires courts to consider four non-exclusive factors: (1) the purpose and character of the use; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used; and (4) the effect upon the potential market for the copyright work. The Supreme Court’s decision in Campbell v. Acuff-Rose Music, Inc. emphasized that transformative use (whether the secondary use “adds something new, with a further purpose or different character”) is central to the first factor analysis. However, its more recent ruling in Andy Warhol Foundation v. Goldsmith clarified that the inquiry must consider whether the new use serves a sufficiently distinct purpose or character or “merely supersede[s] the objects of the original creation,” shifting the emphasis toward market substitution over artistic transformation.

The fourth factor, which assesses market harm, has been referred to as “undoubtedly the single most important element of fair use” according to the Supreme Court in Harper & Row Publishers, Inc. v. Nation Enterprises. Importantly, this factor focuses specifically on “the harm of market substitution” rather than all possible competitive effects.

Disaggregating “Uses” Under Warhol

Both decisions engaged with the Supreme Court’s directive in Warhol that courts must parse multiple uses of copyrighted work individually. In Bartz, Judge Alsup applied this framework to distinguish three discrete uses: (1) training LLMs (fair use); (2) converting purchased print books to digital format (fair use); and (3) building a permanent digital library using pirated sources (not fair use). Judge Alsup concluded that this latter use failed fair use not because of bad faith, but because it violated the principle that intermediate copying must be narrowly tailored to a specific transformative purpose.

In contrast, Judge Chhabria in Kadrey acknowledged this analytical approach but ultimately treated Meta’s actions—downloading and ingesting books—as part of one integrated process directed toward the ultimate and highly transformative purpose of training Llama. This represents a more holistic approach to use analysis, consistent with the Second Circuit’s reasoning in Authors Guild v. Google, Inc. (Google Books), in which the Second Circuit explained that Google’s scanning, indexing, and snippet display were all part of a unified, transformative use that served a different purpose than the original works and therefore qualified as fair use.

First Factor: Transformative Use Analysis

The Human-AI Learning Analogy

Both courts addressed defendants’ analogy between AI training and human learning processes. Judge Alsup enthusiastically endorsed it, writing that requiring compensation “…each time [humans] recall [a work] from memory, each time they later draw upon it when writing new things in new ways would be unthinkable.” He distinguished this from Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence, Inc., where the court found the use non-transformative, reasoning that the AI functioned as a search engine rather than a generative system.

Judge Chhabria took a more measured stance. While he ultimately found Meta’s use highly transformative, he acknowledged that LLMs digest works differently than humans, by detecting and learning statistical patterns through billions of iterative exercises.

Commercial Nature, Bad Faith, and Piracy

The courts also diverged on how commercialism and bad faith should weigh in. Judge Chhabria referenced expert estimates suggesting that Meta could generate $460 billion to $1.4 trillion in revenue from its generative AI efforts over the next decade, yet still concluded that commercialism is less important when the secondary use is highly transformative. On the bad faith issue arising from defendants’ use of “shadow libraries,” he noted conflicting precedents, observing that while Perfect 10, Inc. v. Amazon.com, Inc. suggests good faith requirements, Google LLC v. Oracle America, Inc. expressed skepticism that bad faith has any role in the fair use analysis. In the face of this doctrinal uncertainty, he declined to give bad faith decisive weight.

By contrast, Judge Alsup declined to weigh in on bad faith and instead evaluated the use of pirated books under the intermediate copying doctrine developed in Sega Enterprises Ltd. v. Accolade, Inc. and Sony Computer Entertainment, Inc. v. Connectix Corp. He reasoned that intermediate copying can qualify as fair use only when it is necessary to create a transformative product. Because Anthropic retained pirated books indefinitely for broad, undefined purposes, he found that this failed the Sega test. In effect, he concluded the legality of the source material and the necessity of its use were dispositive. Regardless of the defendants’ subjective intent, the books were illegally obtained and not essential to the specific transformative use. As such, they could not justify fair use.

Third Factor: Amount and Substantiality

Both courts found wholesale copying of copyrighted works reasonable in service of defendants’ transformative purposes. Judge Chhabria stressed that what matters is not the amount of copyrighted material used by the copier in private, but rather the amount of copyrighted material subsequently made publicly available, citing Fox News Network, LLC v. TVEyes. Since neither LLM revealed meaningful portions of plaintiffs’ works in output, this factor supported fair use.  

Fourth Factor: Market Harm Analysis

Direct Substitution Theory

The courts rejected fears that LLMs could directly substitute for original works through regurgitation. Judge Chhabria noted that expert testimony showed Llama could generate at most about 50 words from any plaintiff’s book under adversarial prompting. This fell well below the threshold established in Google Books, where the Second Circuit permitted 16% snippet views without finding harm.

Licensing Market Theory

Judge Chhabria also dismissed the argument that unauthorized AI training harmed a licensing market for training datasets. Applying Tresona Multimedia, LLC v. Burbank High School Vocal Music Association, he held that losing license fees for a transformative use is not cognizable harm. This reflects the Supreme Court’s caution in Oracle against circularity in licensing-based harm theories, where plaintiffs may argue that fair use fails simply because they would have charged a fee for the use.

Market Dilution: The Critical Divergence

One of the key doctrinal splits was how courts looked at potential market dilution that might be caused by AI-generated works that compete with but do not directly copy plaintiffs’ works. Judge Alsup downplayed dilution concerns using a human analogy: someone reading classics and emulating their style does not violate copyright, so neither does an LLM doing the same. He characterized this as ordinary competition that is not in line with the kind of competitive or creative displacement that concerns the Copyright Act. This, he said, reflects permissible competition, not unlawful substitution.

But while Judge Alsup did not appear to entertain market dilution, noting that the Copyright Act seeks “to advance original works of authorship, not to protect authors against competition,” Judge Chhabria gives it more serious consideration, acknowledging the unique threat of AI’s potential to flood the market with endless amounts of images, songs, articles, books, and more. Observing that generative tools can flood markets with infinite outputs requiring very little time or creativity, he warned that this harm could be dispositive in future cases. In a world where generative tools can churn out stories faster than a novelist can finish a sentence, the risk is not duplication—it’s displacement. Nonetheless, he found that the plaintiffs’ record in Kadrey was ultimately too thin to prevail.

Here, Judge Chhabria’s treatment of market dilution as potentially dispositive represents an important doctrinal development in the application of fair use to new technology. His framework suggests that future plaintiffs who can show with specific evidence that generative AI floods their creative markets could defeat fair use defenses, regardless of strong transformative use findings.

Doctrinal Implications and Future Litigation

Use Disaggregation

The courts’ differing applications of Warhol’s “use disaggregation” standard will likely require appellate resolution. Judge Alsup’s compartmentalized analysis offers more predictable outcomes but may not adequately account for integrated technological processes. Judge Chhabria’s holistic approach better reflects modern technological reality but provides less analytical and doctrinal clarity.

Evidence Development Requirements

Both decisions emphasize the critical importance of evidence development in fair use litigation. Judge Chhabria’s grant of summary judgment to Meta, despite acknowledging serious concerns about potential market harm, demonstrates that compelling theoretical arguments alone cannot substitute for a developed factual record.

Conclusion

These decisions establish divergent analytical frameworks with the potential to shape AI copyright litigation for years to come. While both courts reached favorable outcomes for the AI companies, their reasoning suggests that cases with better-developed records on market dilution may yield different results, giving plaintiffs a potential road map for future litigation. The fundamental tension between transformative use doctrine and market substitution analysis remains unresolved, however, leaving much uncertainty.  

The legal terrain is in flux. Though limited in scope, these decisions carry broad implications and signal that copyright disputes over AI training will continue to unfold. As Judge Chhabria warned, even if training an LLM is technically transformative, fair use becomes harder to justify when it relies on copyrighted books to power a system capable of generating massive profits and flooding the market with competing works. The coming years will test how courts balance innovation against authorial rights, and whether copyright law can evolve without eroding the creative incentives it was designed to protect. This post isn’t legal advice—just general insight. Always consult a qualified lawyer about your specific situation. ESCA Legal specializes in intellectual property, global brand strategy, and fighting counterfeits; reach out to us at info@esca.legal if you’ve got questions.