The year 2022 is destined to have its place in the chronicle of AI-related legal disputes since this year, not only did the Copyright Office for the first time grant registration for an AI-assisted artistic creation but the long-expected copyright infringement cases against AI art have finally started to emerge.[1] In the first case the office granted and later revoked the registration for a children's storybook that uses AI assistance for illustrations, but that temporary breach in the wall preventing non-human authorship from obtaining registration gave the industry enough to hope for. In the second, a group of artists has been suing several platforms for using their productions in generative AI algorithms without proper acknowledgment. Whether a tide of suits will follow very much depends on the outcome of this class action.
The family of learning algorithms, to which the AIs in both incidents belong, emulate human artists by processing a large volume of human-created paintings/writings is generally referred to as learning. Accordingly, generative AIs are the AIs that are trained by the learning algorithms to independently produce art or to provide assistance to human artists in the process of art creation. However, from the very beginning of its emerging popularity, the massive commercial use of this kind of AIs raises serious suspicions that such uses may infringe on the copyrights of the artists whose works have been used for training the AIs as their and AI’s products will almost certainly compete in the same market.[2] Indeed, while pieces of art that are purely created by AIs can never be registerable copyrights without substantial human involvement, the other side of the story is far from clear: will artists have the legal standing to sue AIs that are trained on their productions for copyright infringement?[3] While to this day not a single U.S. court has ruled on the issue, reasonable speculations can be and are already being made on how the court will find a balance between the creative values in these AIs and the copyright law.
Generally, two elements of the copyright infringement test in cases short of direct exact copying are relevant to the current discussion: 1) whether the AI production is substantially similar to its human-created attendants and 2) even if the AI production is dangerously similar to the human antecedent, whether it can still be fair use, while other elements, including originality, are assumed to be met given how easy it is to be met and the fact that the involvement of AI does not change the legal conclusions drawable from these discussions.[4]
Substantial Similarity
The most intuitive question to ask in a potential AI infringement case is whether the produced copy is similar enough to the original piece to the point that the artistic expressions like arrangements, structures, languages, and stokes, as opposed to the ideas, styles, and genres of it, can become confusingly similar to the eyes of a reasonable observer.[5] This is an inherently vague and fact-intensive element, but for AI this is inherently trickier since at least part of the algorithm works by probability modeling which, simply put, extracts the most common features/arrangements from the database it is learning from.[6] The pieces generated this way, especially when viewed as a whole, either reflect a common style, common domain techniques, or a mixture of familiar elements used by different authors. This situation creates two difficulties for potential suing artists. First, an artist who only recognizes a similar style/approach to the same theme in the AI piece probably does not have a prima facie case since, as mentioned, the idea behind the piece is not what copyright protects.[7] Second, one who recognizes a copy of the exact angle/rendition/arrangments of colors/elements may still face dilution by elements the AI learned from other pieces given that none of the more detailed elements are dispositive elements; most jurisdictions will consider them as evidence building toward a prima facie case.[8] Moreover, in terms of how detailed the court’s comparisons will be, there exist quite varying levels of flexibility from jurisdiction to jurisdiction and case to case as some courts insist on looking at the production as a whole while others are more ready to delve into artistic details.[9] Whether the court will side with a suing artist whose artistic element has only been appearing in a small portion of the AI production very much depends on whether the court is more willing to look at pieces as a whole or give it a more detailed, element-wise comparison. However, if an AI model has been exclusively trained on the production of a single author, its productions will almost certainly fall under the derivative work rights of the author which cover works that have been based on/derived from the author’s own productions in a recognizable way.[10]
Fair Use
Another potential obstacle for the suing artists is fair use. This element, however, requires separate discussions for training AIs and applying the AIs for commercial purposes. As provided by section 107 of the Copyright Act, whether an otherwise infringing action can be a fair use of the copyrighted work depends on 1) the purpose and character of the use, 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) effect of the use upon the potential market for or value of the copyrighted work.[11] The first element is famously known as a response to a series of First Amendment challenges to copyright law in general and it enshrines many speech interests like parodies and nominative uses in criticisms.[12] This close connection to the First Amendment also means that it is highly likely that a court will find a nonprofit/academic researcher’s use of artistic productions in training an AI to be fair use as academic research/publication has been one of the most important First Amendment interests. Indeed, in a series of cases like Cambridge Univ. Press v. Albert, the courts have consistently found that the first element in the fair use test strongly favors nonprofit/education/research uses while under these circumstances it is less likely that such uses will affect the market and the value of the original productions.[13] However, such observations do not seem to apply to the commercial use of AI models in mass-producing and automating the art creation process. Not only is it usually for profits, but the very scenario where art has become automated has been a nightmare for many artists as their works’ value may evaporate, or at least, be fundamentally altered, overnight.[14] But there are still the possibilities that such AI productions will be ruled as fair use under the Supreme Court case, Campbell v. Acuff-Rose Music, Inc., in which alongside the four factors mentioned, the court will further consider how transformative the new production has been, i.e., whether it creates a totally different and new object, purpose, or expressions based on the first production.[15] This kind of fair use has become particularly important right now since another case is currently pending before the Supreme Court on whether Andy Warhol’s famous pop style which features repetitions of a single photo under different stylistic reprinting is sufficient to be transformative use of the original photo.[16] If the Supreme Court affirms the district court that the stylistic re-rendering of the photos and the juxtaposition of them in an unexpected way is sufficient to deliver a transformative message, then it becomes possible that AI production, the defining feature of which is exactly the kind of re-arrangements that often challenge human imagination, will also be transformative fair use. It is still too early to make such a prediction but the possibility is certainly open. The possibility to apply the doctrine of fair use also provides an early opportunity for the court to test out ways to balance the development of AIs and the value of human labor, a task the courts will face eventually.
In summary, while there cannot be any certain legal conclusion regarding whether an artist whose works have been used to train generative AIs can sue the AIs/their owners, it is a field in which stakes are quickly building up. Interested observers should keep watching for the cases that are being fielded with a special focus on the substantial similarity test and the fair use doctrine as their application to AIs may have different consequences from their more general applications.
[1] See https://www.theverge.com/2022/2/21/22944335/us-copyright-office-reject-ai-generated-art-recent-entrance-to-paradise eventually revoked for further considerations; https://www.theverge.com/2022/11/8/23446821/microsoft-openai-github-copilot-class-action-lawsuit-ai-copyright-violation-training-data.
[2] https://www.wipo.int/wipo_magazine/en/2017/05/article_0003.html
[3] See https://www.copyright.gov/rulings-filings/review-board/docs/a-recent-entrance-to-paradise.pdf both the copyright office and common law copyright requires human authorship.
[4] See https://www.fasthofflawfirm.com/anatomy-copyright-infringement-case-elements-copyright-infringement-claim/
[5] 17 U.S.C. §102; Steinberg v. Columbia Pictures Indus., 663 F. Supp. 706, 712 (S.D.N.Y. 1987)
[6] https://developers.google.com/machine-learning/gan/generative
[7] But see Steinberg v. Columbia Pictures Indus., 663 F. Supp. 706, 712 (S.D.N.Y. 1987) where the court ruled styles of painting can also be an element of expression.
[8] 1 Nimmer on Copyright § 2.14 (2022)
[9] See Roth Greeting Cards v. United Card Co., 429 F.2d 1106, 1110 (9th Cir. 1970) as the court looks at the total concept and feel of the production; Compared to Steinberg v. Columbia Pictures Indus., 663 F. Supp. 706, 712 where the court checked a whole list of artistic details in the background.
[10] For a more detailed analysis see https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4022665.
[11] 17 U.S.C. §107.
[12] Chi. Bd. of Educ. v. Substance, Inc., 354 F.3d 624, 628 (7th Cir. 2003)
[13] Cambridge Univ. Press v. Albert, 906 F.3d 1290, 1294 (11th Cir. 2018)
[14]https://bootcamp.uxdesign.cc/between-fear-and-opportunity-when-ai-generated-art-disrupted-the-world-faedb09e364; see also https://www.theverge.com/23444685/generative-ai-copyright-infringement-legal-fair-use-training-data.
[15] Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 571, 114 S. Ct. 1164, 1167 (1994)
[16] https://www.npr.org/2022/10/12/1127508725/prince-andy-warhol-supreme-court-copyright