Locke Lord QuickStudy: Warhol’s Brush With AI: Redefining Fair ‎Use for Generative AI ‎and Large Language ‎Models‎

Locke Lord LLP
June 27, 2023

The Supreme Court recently revisited the applicability of “fair use” as a defense to copyright claims. While the case involved two artistic titans in their fields—Prince and Andy Warhol—the decision will have enormous effects on the emerging-but-already-colossal AI industry, as owners of these AI models have already asserted that their scraping of copyrighted works is protected by “fair use.” Below, we will preview the implications of this ruling on AI models. (For a more complete comment on the Warhol decision, we recommend Locke Lord LLP’s recent discussion on the blockbuster case.)

One of the primary legal claims made against generative AI models (or more precisely, owners of such models), is for copyright violations. Because large language models indiscriminately scrape the internet for content, these models necessarily ingest works protected by copyright. The subsequent use of copyrighted works to assist in machine learning is, according to copyright owners, a violation of the owners’ copyrights. Apart from using copyrighted works to train AI models, copyright owners are also concerned about AI generating outputs that violate their intellectual property rights.

“Fair use” is a legal defense available to users of others’ copyrights; it permits use of copyrighted material, without permission of the copyright owner, in certain situations such as criticizing, educating, and commenting. This judge-made, and subsequently codified, doctrine is intentionally flexible and “permits courts to avoid rigid application of the copyright statue when it would stifle the very creativity which that law is designed to foster.”[1] In practice, courts frequently find fair use when “non-authorized” users (“secondary users”) “borrow” from a copyrighted work to comment on or criticize the underlying work or to educate others through the work.[2]

Fair use will be a first-line defense for those who operate generative AI models and who use those models’ outputs.

1.  Fair Use Doctrine and the Warhol Holding

To provide stability to this flexible doctrine, Congress enumerated four non-determinative, non-exhaustive factors courts are to weigh when deciding whether use of copyrighted work is “fair”:

(1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;

(2) the nature of the copyrighted work;

(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and

(4) the effect of the use upon the potential market for or value of the copyrighted work.[3]

The Warhol decision exclusively analyzed the first factor, “the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes.” [4] “Transformation” and “transformative uses” are frequently used shorthand for discussing the first factor. The more transformative the subsequent use, the more likely such use favors the first fair use factor.

The central holding of Warhol is best explained by the disagreement between the majority and the dissent regarding a quote from a seminal fair use case, Campbell.[5] In Campbell, the Court defined “transformative” use as one that “alter[s] the first work with new expression, meaning, or message.”[6] Taken literally, this quote strongly implies that courts are to compare the two works’ artistic expression, meaning, and message, to determine whether the later use “transformed” the prior, copyrighted work.

The Warhol majority did not take this quote so literally. To the majority, this language is illustrative of the general rule that the “meaning of a secondary work . . . should be considered to the extent necessary to determine whether the purpose of the use is distinct from the original.”[7] That is to say, new “meaning” of a secondary work is relevant to, but not determinative of, whether a secondary use has a “transformative purpose.” This is a necessary premise to the majority’s central holding: that the purpose of the secondary work is what must be “transformative,” not the expression, meaning, or message. Having established this rule of law, its application to the facts was relatively straightforward. Goldsmith’s (the plaintiff and copyright holder) and the Warhol Foundation’s (the defendant and subsequent user) respective purposes were nearly identical: both were portraits of Prince used in and licensed by magazines to illustrate stories about Prince.[8] Accordingly, because of these nearly identical purposes, and notwithstanding the apparent differences in expression, meaning or message, the majority found that Warhol’s subsequent use of Goldsmith’s copyrighted works did not favor fair use under factor one.

In contrast, the dissent took the Campbell quote by its literal terms. The central question, argues the dissent, should not be whether the secondary work transformed the original’s purpose, but whether the secondary work added a transformative element that altered the original with “new expression, meaning, or message” (irrespective of purpose).[9] Accordingly, the dissent chides the majority as being “uninterested in the distinctiveness and newness of Warhol’s portrait”[10] and continually extols the virtues and transformative meaning of Warhol and the at-issue art piece. This makes sense, as the dissent is concerned with transformative expression, and Warhol is the master of saying much by, seemingly, doing little.

2.  Warhol’s Effect on Generative AI

While one can easily point to the hypothetical troubles in applying the dissent’s approach (judges with lifetime tenure may not be the best arbitrators of what qualifies as “new [artistic] expression”[11]), the difficulties of applying the Court’s holding will have non-hypothetical effects. Emerging technologies, like generative AI, that will seek fair use’s safe harbor for the first time, will be shaped by this doctrine’s application.

The Warhol majority and the dissent take pains to, respectively, narrow or widen what constitutes a “transformative purpose” under the majority’s approach. According to the majority, “transformative purpose” is a necessarily narrow inquiry—one that is judged at the granular level. In determining the parties’ respective purposes, the majority looked at the parties’ respective profit motives and the new “aesthetic” and message of the secondary work.[12] As the majority continually noted of the two at-issue works, “both are portraits of Prince used in magazines to illustrate stories about Prince.”[13]

By narrowing the focus on “transformative purpose” to such a granular level (perhaps in response to the critical dissent), the majority has strengthened many potential claims for “fair use.” If, as cautioned by the dissent, courts were to take a broader view, commercialism would subsume all purposes: the profit is the purpose.[14] However, the majority rebuts this attack by reminding the dissent that the case before them involves both the copyright holder and the secondary user who “sold images of Prince . . . to magazines to illustrate stories about the celebrity.”[15] These identical purposes weigh against a finding of fair use. The majority presents a compelling hypothetical to this point: Warhol’s use of Campbell Soup’s logos in his iconic Soup Cans series. In this scenario, Campbell Soup created the logo to sell soup, while the Soup Cans series uses “the copyrighted work for an artistic commentary on consumerism.”[16] Due to this “transformative purpose,” according to the Warhol majority and fair-use precedent, Warhol’s use of Campbell Soup’s logo would be “fair.”[17]

The results of this “transformative purpose” analysis will turn on how far a court zooms in to each parties’ respective goals. To use the Warhol example, was the purpose of the two Prince prints: corporate profit >>> to increase magazine sales, generally >>> or to increase magazine sales of a particular issue about Prince by using a particular image of Prince? The Warhol court answered that question on the right end of this spectrum. As the majority stated, “the analysis here may be different Orange Prince appeared in an art magazine alongside an article ‎about Warhol” (as opposed to an article about Prince).[18] ‎Courts, following this precedent, should take a similarly granular approach.

It should also be noted that this “transformative purpose” analysis appears to presuppose the availability of a market to license the original, copyrighted work. The Warhol court continually stressed that the underlying, copyrighted work was licensed to magazines: the same “purpose” of the defendant. But what if the underlying copyright holder never desired to license the Prince photograph in the first place, would a subsequent use necessarily be transformative? What if there was no market to license this work, would the copyright holder be out of luck? The Court does not provide a clear answer to these questions.[19] In the digesting and learning context, owners of generative AI and large language models should get some comfort out of the Warhol decision. Generative AI models use copyrighted works (along with many more non-copyrighted works) to train a machine on vast amounts of data. This data ingestion allows the machine to generate novel responses to queries by predicting the next word, line of code, etc. It is fair to say this purpose is distinct from that of nearly every copyrighted work in history. This analysis will necessarily change, however, as large language models become more prevalent, and these models become tailored for specific purposes.

A thornier question involves the outputs of generative AI, and the numerous purposes to which people will put them. These “end user” purposes will be numerous and distinct. For example, if you ask for a computer code with specific parameters and the AI generates a response that draws heavily from a copyrighted work, would the purposes be the same? The answer should change based on each parties’ respective goals and the medium in which the copyrighted work was used. The issues with this analysis are compounded by the difficulty of connecting any specific copyrighted work with a specific output. (In a recent complaint, Getty Images was able to overcome this difficulty by pointing to the smeared “Getty Images” watermarks on various outputs created by generative AI.) Furthermore, owners of generative AI models could be exposed to secondary liability for certain infringing outputs, particularly if the owners do not enact guardrails to prevent an AI model from creating works that clearly violate copyright.

3.  The Fair Use Escape Valve

As mentioned above, fair use is an “equitable rule of reason” that “permits courts to avoid rigid application of the copyright statute when, on occasion, it would stifle the very creativity which that law is designed to foster.”[20] Consistent with these general principals, the application of fair use “requires judicial balancing, depending upon relevant circumstances, including ‘significant changes in technology.’”[21]

For more than 200 years, courts have recognized copyright and its limitations, which together protect against “two extremes equally prejudicial”: on the one hand, creative work is to be encouraged and rewarded; on the other hand “the world may not be deprived of improvements, nor the progress of the arts be retarded.”[22] Fair use attempts to strike this balance. However, particularly in times of technological upheaval, courts and practitioners will be left to determine where between those two competing purposes is the right place to draw that line.


[1]Stewart v. Abend, 495 U.S. 207, 236 (1990) (internal quotations marks omitted).

[2] See 17 U.S.C. § 107 (listing examples of fair use).

[3] Id.

[4] Andy Warhol Foundation for the Visual Arts, Inc., v. Goldsmith et al., 598 U.S. __ (2023) (slip op., at 2) (“Warhol”).

[5] Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994).

[6] Id. at 579 (1994)‎

[7] Warhol (slip op., at 32) (emphasis added).

[8] Warhol (slip op., at 13, n.5).

[9] Warhol (Kagan, J., dissenting) (slip op., at 14)

[10] Id. at 1.

[11] Id. at 2.

[12] Warhol (slip op., at 29).

[13] Id. at 22-23.

[14] Warhol (Kagan, J., dissenting) (slip op., at 12-13, n.5).

[15] Warhol (slip op., at 29, n.15).

[16] Id. at 27.

[17] Id.

[18] Id. at 24, n.12.

[19] While the Court does not address question, others have touched upon it. See Nimmer on Copyright § 13.05 (“If the defendant’s work adversely affects the value of any of the rights in the copyrighted work . . . the use is not fair, even if plaintiff has not yet exercised that right.”).

[20] Stewart, 495 U.S. at 236.

[21] Google LLC v. Oracle Am., Inc., 141 S. Ct. 1183, 1197 (2021) (emphasis added); see also, Twentieth Century Music Corp. v. Aiken, 422 U.S. 151, 156 (1975)‎ ‎(“When technological change has rendered its literal terms ambiguous, the Copyright Act must ‎be construed in the light of its basic purpose[, to ‘stimulate artistic creativity’].”)‎

[22] Sayre v. Moore, (1785) 102 Eng. Rep. 138, 139 n.(b).

Click here to visit the AI resource center