Fair use benefits everyone

The Court should reject Plaintiffs’ attempt to extend copyright beyond its established limits. AI training uses copyrighted works for a fundamentally different purpose, does not provide the public with substitute access to those works, and does not cause the market harm cognizable under the Copyright Act. Instead, it enables a broad array of transformative technologies that are already delivering public benefits. Copyright exists to promote creativity and progress, not to suppress lawful innovation or protect authors from competition. Settled fair use principles therefore compel the conclusion that generative AI training is fair use.

This case concerns more than companies that develop large language models (LLMs). It affects medical researchers and patients, educators serving students with different learning needs, people with disabilities seeking greater independence, farmers applying agricultural knowledge, independent creators developing new forms of expression, and industry sectors from software to the law. These communities share an interest in the ability to study and learn from broad and diverse sources of knowledge, and in access to the tools that make that learning useful. Many depend on systems developed by others, including open source AI systems, rather than training models themselves. Fair use supports their ability to learn, discover, communicate, and create. Restrictions on fair use training would affect both the institutions building AI and the people using it.

The market dilution theory asks copyright law to protect rightsholders against an economic phenomenon that it has never restricted before. New works can reduce the sales, visibility, or value of older works without copying their protected expression. The Copyright Act deliberately leaves that competition outside the copyright owner’s control. Amicus therefore respectfully asks the Court to hold that “market dilution,” i.e., aggregate competitive pressure from noninfringing outputs, including those that occupy the same subject matter or genre, that share unprotected characteristics, or compete for the same audience, is not cognizable market harm under 17 U.S.C. § 107(4). Rejecting the theory would preserve the distinction the Second Circuit has always maintained: copyright protects authors against substitution for their protected expression, not against the success of new expression that merely competes with it.

The modern creative industry’s experience with generative AI provides an excellent illustration of these principles in action, demonstrating how generative AI tools advance copyright’s central purpose to enable further human creativity. For example, musicians use generative AI to explore arrangements, develop melodies, and bring projects within reach that limited resources might otherwise foreclose. These tools support and help to effectuate the artist’s judgment, expanding the possibilities for human expression and the works available to the public.

[T]raining a generative AI model is among the most clearly transformative uses a court has been asked to evaluate. The copyrighted works are not copied to be consumed or distributed; they are analyzed to extract statistical patterns that enable a tool to generate new text or other output at the direction of a human user. Both the purpose and the character of this use are categorically different from the original purpose and character of the underlying works.

Professors of Law

These noninfringing outputs are not substantially similar to Plaintiffs’ works but are simply in the same genre or contain similar facts. Plaintiff’s unprecedented theory of “dilution” turns copyright on its head. Genres and facts are uncopyrightable. Under the Copyright Clause, these elements are not original “writings” of any “author” and are free to copy.

Plaintiffs are asking the Court to throw that caution to the winds and rewrite copyright law to include a lawful acquisition requirement that finds no basis in statute or case law; and to tilt the fair use analysis against the developer of a general-purpose tool if that tool might be used to create potentially competing non-infringing works. The Court should reject both theories.

Startups’ use of foundation models demonstrates that using copyrighted works to train an AI model is a fair use, not copyright infringement. The purpose of training a foundation model is not to reproduce the works it learns; it is to build a general-purpose tool that others can adapt to new ends. The ends startups have chosen bear no resemblance to the purposes of the works in any training dataset.

To the extent AI training enables the creation of new knowledge, expression, software, and other capabilities, those benefits bear on factor four’s assessment of the use’s market effects. That reasoning supports a novel application here: considering the national-security value of capabilities that enable new analysis, software, and other productive work.

Library technology expert

“[T]he primary objective of copyright is not to reward the labor of authors, but to promote the Progress of Science and useful Arts.” Libraries share this fundamental purpose of the Copyright Act. For example, the Library of Congress states its mission is to provide “access to a rich, diverse and enduring source of knowledge to inform, inspire and engage you in your intellectual and creative endeavors.” Across other institutions, the language is remarkably consistent: Harvard Library’s mission is to “engage with our communities in the creation and sharing of new knowledge,” MIT Libraries seek to “advance knowledge by providing a trusted foundation for the generation, dissemination, use, creative engagement with, and preservation of information” and the New York Public Library exists to “advance knowledge by providing free and open access to materials.” The open availability of general-purpose AI models serves and supports an important public benefit in that it enables libraries to keep pace with the preservation and accessibility of human knowledge for future generations.

Professor of Strategic Management

Copyright supports not only the incentive to create content but also the incentive to build tools that learn from that content, and a rule that treated every displaced sale as actionable can weaken the second without reliably protecting the first.

Author

AI is a transformative technology that accelerates the creation of novel works across countless disciplines in the arts and sciences. Mr. Kelly has experience in how AI advances the field of literature. In his view, each AI model can be understood as an entirely new medium that offers authors the capability to create new texts and ideas, and to reason. A ruling for Plaintiffs on fair use would therefore grant Plaintiffs far more than the limited monopoly to their original expression Congress intended to confer. It would hinder the development of technologies capable of disseminating previously unimaginable quantities of information. It would also impede new modes of thought and the creation of new original works in both the arts and sciences. Amicus files this brief in support of Defendants’ argument that the use of copyrighted works to train large language models is a fair use under 17 U.S.C. § 107. Based on his lifelong pursuits of writing, the protections afforded to copyright owners are important to Mr. Kelly. Equally vital is preserving room for permissible uses of copyrighted material and the growth of transformative technologies. Otherwise, copyright will stifle the very creativity, innovation, and spread of knowledge it exists to encourage. The fair-use doctrine is critical to ensuring that balance.

Fair use exists precisely to prevent copyright law from being used against the innovation it was designed to advance. To fulfill copyright’s constitutional mandate “[t]o promote the Progress of Science and useful Arts,” U.S. Const. art. I, § 8, cl. 8, the Court should hold that the use of copyrighted works in training OpenAI’s large language models constitutes fair use as a matter of law.

Professor of Law

On the record in this case, the defendants’ use of plaintiffs’ works for AI training is highly transformative and strongly favors a finding of fair use because under realistic conditions the outputs of defendants’ models are in no relevant way similar to the invisible intermediate copies made during the training process.

Musician and composer / comedian and writer

This brief offers a window into what that sentence means in practice through the work of two independent creators who rely on copyright protection and AI tools to sustain their livelihoods. Amici Sarah Rose Siskind and Michael Wall are artists who create and license original works. But they use AI to handle work surrounding the creative process and to grow their businesses in a market teeming with better-resourced competitors. For Amici, AI does not displace creativity or shortcut their creative processes. AI helps them create more, and to extend their work into new modes of expression. Amici’s experience underscores the United States’ view that AI tools “can and should help level the playing field between mainstream and independent publishers.” Siskind uses AI in every aspect of her business, from day-to-day administrative tasks to developing ideas and bringing in new clients. With AI, her noncreative work is faster and less expensive, and her creative work is better. As she tells it, “I don’t think I could run my business without AI at this point.” Similarly, AI helps Wall manage the operational demands of his music business, reduce costs, and devote more of his time to composing music. Like Siskind, he doubts that his business could survive in its current form without AI.

Computer Scientists

Generalization and memorization are distinct phenomena. Training on a large and varied collection enables a model to respond to prompts not encountered during training, supporting the conclusion that it learns patterns applicable to new contexts rather than simply storing source works. A trained model is not an addressable document repository and does not ordinarily retrieve identified training data as stored files. There is no address, index, or file handle by which particular training data can be requested and returned; generation instead proceeds by producing a probability distribution for the next token and selecting from it.

Fair use has protected tools that help people navigate the vast collections of information the digital world has brought us, including image search and full-text book search. See Perfect 10, Inc. v. Amazon.com, Inc., 508 F.3d 1146, 1164–68 (9th Cir. 2007); Authors Guild v. Google, Inc., 804 F.3d 202, 216–17, 225 (2d Cir. 2015) (“Google Books”). It has also protected libraries’ provision of accessible digital books to readers with print disabilities. See Authors Guild, Inc. v. HathiTrust, 755 F.3d 87, 101–03 (2d Cir. 2014). And it has protected free online publication of technical standards incorporated into law, recognizing the “substantial public benefits of free and easy access to the law.” Am. Soc’y for Testing & Materials v. Public.Resource.Org, Inc., 82 F.4th 1262, 1271–72 (D.C. Cir. 2023). Fair use will be equally beneficial in ensuring that AI’s benefits accrue to the public at large, instead of being controlled by a privileged few.

Former members of Congress

“From its beginning, the law of copyright has developed in response to significant changes in technology.” Sony Corp. of Am. v. Universal City Studios, Inc., 464 U.S. 417, 430 (1984). Consistent with that provenance, the doctrine of fair use, which was made statutory law through the Copyright Act of 1976, is meant to work in tandem with technological change, not against it. AI is the latest and the most consequential technology to which that doctrine must be applied. The use of copyrighted works to train AI models is a paradigmatic case of fair use. An AI model relies on works not to communicate their expression to the public but rather, quite differently, to derive the statistical patterns of language that allow the model to function. A ruling that AI training constitutes a fair use is vitally important to our Nation. The freedom of American scientists, companies, and entrepreneurs to train AI is of paramount economic and strategic importance.

Pulitzer Prize-winning journalist / Professor of Law

OpenAI’s use of newspaper articles to train its generative AI models didn’t harm the reasonable commercial expectations of the News Plaintiffs in any way. They commissioned and published newspaper articles with the expectation of selling consumers access to those articles. They retain and continue to exercise the exclusive right to sell consumers access to their articles. Their ability to sell consumers access to their articles wasn’t affected by OpenAI’s use of those articles to train its AI models.

In this litigation, the New York Times seeks to narrow fair-use doctrine to exclude the training of OpenAI’s large language models (LLMs). That result would be inconsistent with basic copyright law principles and severely hamper “the Progress of Science and useful Arts.” U.S. CONST. art. I, § 8, cl. 8. As discussed below, publications—including independent and start-up media, in addition to the New York Times itself—are leveraging LLMs for journalistic and other purposes. And beyond the subject matter of this litigation, LLMs are already helping researchers across fields achieve major breakthroughs. Constraining LLM development under a misunderstanding of fair use doctrine would thwart such creative and scientific progress while hindering American prosperity and economic mobility.