Copyright Liability for Generative AI Pivots on Fair Use Doctrine

On Aug. 24, the Copyright Office issued a request for comment on ways that generative AI technology that is “capable of producing outputs such as text, images, video, or audio” will affect “the future of creative industries” and “the copyright system” as a whole.

It’s not yet clear how much change will be caused by computer systems that are capable of replicating creative processes long considered to be fundamentally human. But some of the legal questions surrounding generative AI are starting to come into focus.

A class of authors and creators including comedian Sarah Silverman has already filed suit against generative AI trailblazers, asserting that the act of including preexisting copyright-protected materials in datasets used to train generative AI systems infringes their exclusive rights under the Copyright Act. These cases involve comedians like Sarah Silverman and, in a case filed Sept. 19, “Game of Thrones” author George R.R. Martin. Fortunately, case law provides important clues as to the framework that will be used to analyze this question.

The 2015 Authors Guild Inc. v. Google, Inc. case provides a good starting point. In Authors Guild, the Second Circuit Court of Appeals held that Google’s practice of digitizing books without authorization to create a searchable database was a “fair use” that couldn’t give rise to liability.

Importantly, the court focused on the “highly transformative purpose” of Google’s use of this material—that Google was transforming many diverse copyright-protected books into a new, useful search mechanism that didn’t compete with the underlying works.

The US Supreme Court’s 2023 decision in Andy Warhol Foundation v. Goldsmith demonstrates an important limit to that defense. There, the court concluded that artist Andy Warhol’s replication of someone else’s photograph in one of his works wasn’t protected under the fair use doctrine. In contrast to the database at issue in the Authors Guild case, the court found that Warhol’s recoloring and replication of the preexisting photo did have the potential to act as a substitute for the original work.

The Supreme Court urged lower courts to focus on the fourth of four statutory fair use factors, advising that the effect that the allegedly infringing work may have on the market for, or value of, the original work should weigh heavily in the analysis.

For generative AI, there are two distinct aspects to this fair use analysis: whether these systems infringe at the outset by copying and incorporating copyright-protected material as part of the learning process, and whether generative AI infringes rights in those same materials in the resulting output that is produced.

First, copyright owners will surely argue that incorporating copyright-protected materials into a data set to be used by a generative AI system without authorization at least facially violates one of the most fundamental rights created by the Copyright Act—the right to control reproduction.

Given the Supreme Court’s recent guidance, the question of whether that duplication is “fair” will likely turn on the factual question of whether there is a real-world market for the right to include works in a generative AI data set, and whether the operator of that AI system effectively competes with the owners of the reproduced works.

Copyright owners have already started to develop this argument. For example, photo licensor Getty Images recently sued Stability AI in Delaware, alleging its generative AI system “competes directly” with Getty’s licensing marketplace for “those seeking creative imagery” including in generative AI datasets. The court’s analysis of whether that allegation is true may provide valuable insights into the standards courts will apply in evaluating market effects.

Second, the question of whether the output from generative AI systems infringes a copyright is likely to turn on the traditional question of whether the AI-generated output is “substantially similar” to the underlying work. Generative AI outputs, of course, are not uniform. Some systems produce short textual answers and others produce complex commercial works. To complicate matters further, generative AI outputs from the same system can change over time, as the model ingests new training data.

At the extremes, the analysis is likely to be easy. To the extent a generative AI system simply replicates the valuable part of a copyright-protected work in ways that can serve as a substitute for the original work, it’s more likely to be found infringing. On the other end of the spectrum, if a generative AI model produces a new and original work that’s unrecognizable as a reproduction, then the risk that the work will be found infringing decreases.

An important part of the analysis will be how these systems imitate style—do they simply regurgitate relevant portions of the underlying works, or do they actually generate new outputs without using meaningful elements from the originals?

The Copyright Act, even as it has been amended over the years, predates generative AI systems that replicate human art in a matter of seconds, and the text isn’t directed to some of the ways an AI system may imitate an original work. For example, the Act doesn’t list abstract concepts like “voice” or “style” as among the bucket of exclusive rights conveyed to creators.

Instead, parody of the voice or style of a creator is well within the heartland of fair use. When a voice or a style becomes so commonly copied that it becomes a genre standard, it’s shielded by doctrines that protect common tropes such as detective stories that feature trench-coat wearing gumshoes.

Decisions in generative AI copyright cases will also have important policy ramifications. Some courts may be hesitant to impose liability that could hinder the development of a nascent, potentially transformative industry. And foreign competitors in the AI space may not have the same concerns for intellectual property that are embodied in US law. Policymakers may be required to step in and help courts to balance these interests.

Only one thing is certain—the answers to these questions will be driven heavily by the effects that these systems have on the market for copyright-protected works.

The cases are Authors Guild v. Google, Inc., 804 F.3d 202, 10/16/15 and Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 143 S. Ct. 1258, 05/18/23.

This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law and Bloomberg Tax, or its owners.

Author Information

Howard S. Hogan is a partner at Gibson, Dunn & Crutcher and is chair of the fashion, retail, and consumer products group.

Connor Sullivan is a senior litigation associate at Gibson, Dunn & Crutcher.

Jeffrey Myers is an associate at Gibson, Dunn & Crutcher.

Write for Us: Author Guidelines