Books, Bots, and the Bottom Line: Reconsidering Copyright Fair Use in Generative AI Training
Credit: Stock Images
In July 2026, a federal court approved a $1.5 billion settlement in a class action lawsuit brought by authors who alleged that Anthropic infringed their copyrights by downloading and storing unauthorized copies of their books. This settlement, aside from being the largest known copyright case settlement in the United States, foregrounded an unresolved question at the center of modern copyright law: what constitutes fair use of protected works in training commercially valuable generative AI models?
Generative AI models are developed through large-scale reproduction of real-world content. To identify linguistic, visual, and semantic patterns, developers ingest vast collections of books, articles, and images. The resulting models typically do not function as archives that regurgitate this data to users. Rather, they apply patterns extracted during training to contextualize interactions and provide informed responses to human prompts. As model development has accelerated, developers have sought larger, more varied datasets to improve the accuracy and authenticity of their LLMs' outputs.
AI companies contend that the use of this media for the purpose of LLM training is transformative, and thus fair use of potentially copyrighted material. Conversely, creators, publishers, and other rights holders express concerns that such use may undermine the economic value of their work by enabling competing services and reducing potential licensing revenue. The central question is whether copying protected works to train a commercial generative AI model is a fair use when the model is not designed to distribute the works verbatim but may nevertheless compete with them or affect an emerging licensing market.
This dispute is governed by 17 U.S.C. § 107, which codifies the fair use doctrine. Fair use provides a defense to copyright infringement claims by permitting certain unauthorized uses of copyrighted works. These exceptions typically apply to criticism, commentary, teaching, and research, and often serve the purpose of promoting innovation and education without free-riding on others’ creative investment.
Section 107 establishes four factors that courts consider when evaluating whether a particular use qualifies as fair use.
First, courts are directed to consider the purpose and character of the usage, and identify whether the work is used transformatively for an entirely different purpose, meaning, message, or function.
Second, the courts review the nature of the copyrighted work. Typically, highly creative or unpublished materials are afforded more protection than purely factual materials which are more likely to have a functional application.
Third, courts consider the amount and substantiality of the material used, including whether the copied portion captures the heart of the original work.
Finally, courts evaluate the effect such usage would have on the existing markets for the copyrighted content, essentially ensuring that the use would not operate as a substitute.
These questions, though, long predate the rise of generative AI. The Supreme Court ruling on Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith provides a useful framework in evaluating them. Following Prince's death, the Andy Warhol Foundation licensed to Condé Nast an image Warhol had created from a photograph by Lynn Goldsmith. Goldsmith argued that the commercial licensing of the Warhol image infringed her copyright as the original creator of the media, and sued the Andy Warhol Foundation. Although Warhol significantly altered the original photograph, the Supreme Court held that the first fair use factor weighed against the Foundation because its commercial licensing served substantially the same purpose as Goldsmith's photograph. Warhol therefore directs courts to evaluate fair use with attention to the defendant’s particular commercial use, rather than treating alteration, technological novelty, or the creation of a new product as independently dispositive.
This ruling helps explain why full-text copying was permissible in Authors Guild v. Google. Here, Google partnered with libraries to scan millions of published books absent permission from copyright holders and used the information to create a searchable database that allowed users to read short text snippets, search books by keyword or phrase, and view relevant bibliographic information. A collective of authors challenged the scanning, storage, and display of their completed works as a violation of their copyrights. The Second Circuit held that Google’s conduct was fair as it was substantially transformative in nature, enabling an entirely new function. Because the platform allowed users to search for specific content without providing a practical means of reading entire books, it did not serve as a substitute for purchasing or reading the original works. The decision established that the fair use statute permits full-text copying where it supports a distinct, non-substitutive, and analytical use.
The same focus on function and market effect led to a different outcome in Hachette Book Group, Inc. v. Internet Archive. There, the Internet Archive scanned physical books and distributed digital copies through its Controlled Digital Lending program. During the COVID-19 pandemic, its National Emergency Library expanded simultaneous access, prompting publishers to sue. Both cases accept that copying an entire work can sometimes be necessary for a transformative purpose; they reach different results because of the ultimate function of their services. Although both Google and the Internet Archive scanned physical books, the latter provided users with full-length digital copies that could substitute for licensed e-books. Its lending model therefore served essentially the same purpose as conventional book lending and failed to qualify as transformative use. As such, this case illustrates that even where complete copying is necessary, fair use depends on the market function of the product, clarifying the validity of complete copying based on its final commercial function.
These principles now frame the recent decisions more narrowly addressing LLM development. In Bartz v. Anthropic, authors asserted that Anthropic had infringed their copyrights by copying and using books in the development of Claude and related LLMs. Their claims concerned several practices, though most importantly the training of LLMs on copyrighted books and the retention of books from pirated sources. The district court held that using copyrighted books to train LLMs constituted fair use and separately found that digitizing purchased print books was fair use. However, it rejected fair use for Anthropic's acquisition and retention of pirated copies in its general-purpose library. This decision aligns with the functional inquiry of Google Books, while affirming the nuance of Hachette and Warhol: every upstream act of copying cannot be justified as fair use via mere attachment to technologically novel products.
Conversely, Thomson Reuters v. ROSS Intelligence marks the opposing end of the emerging doctrine. Thomson Reuters operates Westlaw, a legal-research platform whose editor-written headnotes summarize judicial opinions. ROSS Intelligence, a legal-technology startup, sought to develop an AI-assisted legal research platform by hiring LegalEase Solutions to produce training questions derived from Westlaw headnotes after Thomson Reuters declined to license its materials to them. Thomson Reuters sued for copyright infringement and the Third Circuit upheld this finding, noting that the use of Westlaw headnotes was highly commercial and minimally transformative. The market harm posed by ROSS’s use of the headnotes to develop a competing legal-research platform fulfilling a similar niche as Westlaw furthered the case against fair use. This case, described by the Third Circuit as an ordinary copyright case, does not drastically alter the legal landscape on this matter, but it affirms a core tenet of earlier cases: the first and fourth fair-use factors weigh especially strongly against fair use when protected content is used to develop a product that occupies a similar market space as the original content.
These emerging decisions on AI development seem to require a more nuanced, context-specific application of the first and fourth fair use factors. Generative AI complicates the traditional fair use doctrine as it separates the act of ingesting copyrighted content from the generation of new content. Because models may rely on full-text copying without displaying protected works verbatim, courts must look beyond obvious reproduction when evaluating whether the novel, transformed product poses a harm to the relevant market. Read together, Google Books, Hachette, Bartz, and ROSS do not establish competing rules about full-text copying. Rather, they demonstrate that copying may be fair when it supports a distinct analytical and non-substitutive function, but not when the copying facilitates ordinary consumption of the protected work, supports a substantially overlapping commercial product, or rests on independently unlawful acquisition and retention. To better distinguish innovation from exploitation, courts should apply the fair use statute with close attention to the specific commercial function of the AI service.
First, courts should evaluate transformativeness by examining the defendant’s specific use of the copyrighted material, consistent with Warhol v. Goldsmith. Rather than focusing solely on whether a model produces new content, courts should examine whether the resulting service serves substantially the same commercial purpose as the original work.
Second, courts should not treat the absence of materially similar outputs as conclusive evidence that a generative model has no potential to cause market harm. Although the market effects discussed in the fourth fair use factor often manifest themselves as direct and obvious harms, generative AI may also create subtler pressures by diluting demand for human-created work, depressing prices, or reducing licensing opportunities in creative markets. In the future, courts should accordingly consider whether a real market exists for licensing copyrighted materials for AI training; a proposed licensing right, such as LearnRights, could provide a mechanism for creator compensation. The development of such licensing markets would not necessarily preclude fair use, but could provide evidence of potential market harm under the fourth factor. This is consistent with ROSS, which recognized evidence that AI-training licenses for Westlaw were already becoming commercially viable.
However, courts should not treat factor four as dispositive merely because market harm is difficult to measure. When the use is minimally transformative, and relies on the full-copying of highly creative work, courts should give substantial weight to the first three factors. This approach accords with the broader principle that technological or commercial novelty cannot, by itself, justify extensive copying. These precedents confirm that neither commercial novelty nor complete copying is legally neutral under fair-use analysis. Accordingly, when an AI developer’s use is only minimally transformative, uncertainty surrounding market harm should not be reason enough to uphold a fair use defense. Ultimately, to determine whether an AI developer’s use of copyrighted material is legally permissible as fair use, courts should apply the first and fourth fair-use factors—the defendant’s purpose of use and the use’s effect on relevant markets—by examining whether an AI developer’s specific commercial use is genuinely distinct, and non-substitutive.
Anya Gopalakrishnan is a freshman from California studying International/Public Affairs and Economics. She is a Junior Blog Writer with the Brown Undergraduate Law Review and can be contacted at anya_gopalakrishnan@brown.edu.
Nicolas Abril is a senior from Florida studying Applied Mathematics and Economics. He is a Junior Blog Writer with the Brown Undergraduate Law Review and can be contacted at nicolas_abril@brown.edu.
Hayden Hradek is a sophomore studying International and Public Affairs and Social Analysis and Research at Brown University. He is a Senior Blog Editor with the Brown Undergraduate Law Review and can be reached at hayden_hradek@brown.edu.