ANALYSIS: Generative AI to Test the Boundaries of Fair Use

“This copyright dispute tests the boundaries of fair use,” Judge Pierre N. Leval wrote at the beginning of the Second Circuit’s 2015 opinion in Authors Guild v. Google, Inc. In this case, the author-plaintiffs sued Google for copyright infringement for scanning digital copies of books, without permission from the copyright holders, and establishing a publicly available search function. The court ultimately decided on summary judgment that Google’s copying was fair use under the doctrine by that name.

In 2015, it was inconceivable that those boundaries would be tested again by the creation of generative artificial intelligence and large language models. Now, as we head into 2024, there’s no doubt that generative AI tech companies will seek refuge in the fair use doctrine, as several plaintiffs allege that OpenAI LLC and Meta Platforms Inc. ingest copyrighted books without permission in order to train their large language models.

It’s unlikely that courts will unravel the intricate, cutting-edge development of generative AI models in support of copyrighted works. Rather, courts will rely on the fair use doctrine to allow the transformative use of copyrighted works to train generative AI models for public enrichment.

The Fair Use Doctrine

Authors have an exclusive right under the US Constitution to their respective writings, for a limited time period, in order “to promote the progress of science and useful arts.”

Understandably, US copyright laws were intended to encourage authors to create and release for public consumption “informative, intellectually enriching work,” in exchange for a limited monopoly over their own work. Many courts, however, recognize that the primary and ultimate beneficiary of copyright laws is the public, who consume copyrighted works in order to expand their knowledge and understanding of the world.

In furtherance of public enrichment, Congress enacted the 1976 Copyright Act, which codified the doctrine of fair use, allowing the copying and use of copyrighted work in certain circumstances. Fewer than 10 years later, the US Supreme Court took up the issue of fair use for the first time in Sony Corp. of Am. v. Universal City Studios, Inc., where the court found that the use of video-tape recorders at home to record material broadcasted on television was a fair use of copyrighted works and didn’t constitute copyright infringement.

In 1994, the Supreme Court elaborated on the fair use doctrine in Campbell v. Acuff-Rose Music, Inc., where it first considered “whether and to what extent the new work is transformative.” A transformative work is one that adds something new, with further purpose, the court said.

It was this very reasoning that led the Supreme Court to reject the fair use argument most recently in Andy Warhol Found. for the Visual Arts Inc. v. Goldsmith. Affirming the Second Circuit’s decision in favor of the plaintiff, photographer Lynn Goldsmith, the court recognized that whether a secondary work is transformative depends on “whether it is in service of a fundamentally different and new artistic purpose and character.” The Warhol painting shared substantially the same purpose and function as the original photograph by Goldsmith. They were both portraits of the same person, one imposing its style over another, without more, the court said.

However, the fair use analysis doesn’t stop at “the purpose and character of the use.” Courts evaluate three other factors, with varying degrees of importance. The fourth factor is the most important, and considers “the effect of the use upon the potential market for or value of the copyrighted work.” Does the secondary use reduce the market or value of the original work and its potential derivatives? Is the secondary use “unrestricted and widespread”?

The first and fourth factors are undeniably intertwined. The more transformative the use, the less of an adverse impact it will have on the market, as compared to a work that supersedes the original. In addition, courts recognize that only “traditional, reasonable, or likely to be developed markets” can be considered in connection with potential licensing revenue under the fourth factor.

A Groundbreaking Decision to Come

There’s currently no fair use case that explores the abundance of information consumed by large language models. Any decision rendered on the issue will have far-reaching consequences on a wide range of industries, either crippling the advancement of generative AI or diminishing the exclusivity of copyright.

A strong argument can be made that OpenAI’s and Meta’s use of the copyrighted works is highly transformative and doesn’t usurp the original copyrighted work. The copyrighted works, along with a large volume of other data, are used to train large language models on word usage, semantics, grammar, and sentence structure.

The statistical data derived from these works serves an entirely different purpose than the copyrighted works themselves. It allows the models, for example, to “generate creative text, solve mathematical theorems, predict protein structures, answer reading comprehension questions, and more.” These capabilities are quite distinguishable from the original work.

One of the reasons for OpenAI’s and Meta’s ingestion of the original copyrighted works is to make available significant information about those works (and others) that is otherwise not available—similar to what Google did in Author’s Guild. The copyrighted works are also not being used in the manner originally intended. Moreover, even if the copyright holders establish a market for licensing their works for ingestion by large language models, there’s no evidence that this is a traditional market. Copyright holders can’t prevent fair use by creating or licensing a market for that very use.

This may be an unpopular opinion, but one that courts hearing the generative AI copyright infringement claims are likely to make in the year to come.

Access additional analyses from our Bloomberg Law 2024 series here, covering trends in Litigation, Transactions & Contracts, Artificial Intelligence, Regulatory & Compliance, and the Practice of Law.

Bloomberg Law subscribers can find related content on our AI Legal Issues Toolkit and our In Focus: Artificial Intelligence page.

If you’re reading this on the Bloomberg Terminal, please run BLAW OUT <GO> in order to access the hyperlinked content, or click here to view the web version of this article.