Photo Credit: https://futureoflife.org/cause-area/artificial-intelligence/ (last visited May 19, 2023).

Authored By: Ben Byard

July 7, 2023

Progress of artificial intelligence (AI) is on the rise and shows no signs of stopping. Recently this technology has exploded in popularity with the likes of ChatGPT and art that is created entirely by AI. These feats can only be achieved after the AI has been properly “trained” to achieve a desired outcome.[i] For many AI models, this training involves taking in many outside images or sources to be able to adequately “learn” what it is trying to do.[ii] This training usually requires thousands or tens of thousands of outside materials to be given to the AI for it to function correctly.[iii] Since most people or companies do not own the requisite amount of images or data themselves, they must use images or data from outside sources.[iv] Checking for copyrights and protections on all of these images would be both extremely time-consuming and expensive, so nearly all AI companies have used copyrighted images to train their AIs without worrying about the potential consequences. This has led to new developments in copyright litigation, where it is unclear whether using copyrighted images to train AI would violate the copyright.

            A very recent case that is currently undecided will provide some of the first laws directly relating to this topic.[v] Getty Images brought a complaint against Stability AI for infringing on their copyright in twelve million images that were used to train Stability’s AI.[vi] While it will take some time for this case to finish, previous case law can still provide some guidance on whether AI would infringe on other’s copyrights. While Stability AI has only responded on jurisdictional grounds to Getty Images, if these arguments fail it is most likely they will make an argument for the fair use of the copyright.[vii] The US Copyright Act allows the use of copyrighted material through the fair use doctrine.[viii] This allows a copyrighted material to be used or reproduced without punishment if the copyrighted work is used “for purposes such as criticism, comment, news reporting, teaching (including multiples copies for classroom use), scholarship, or research. . . .”[ix] Campbell v. Acuff-Rose Music, Inc. also clarifies that the more transformative a work is, the less the other factors against fair use will weigh.[x] This case states that transformative works take the original and add some new value or meaning to it, giving parodies as an example of transformative work.[xi]

This would be the most likely argument for the works produced by AI not to infringe on copyright. AI generators take in thousands upon thousands of images and use those to create a new image based on what the AI has “learned.” Since a single image cannot be used to train an AI completely, it would be hard to argue that an entirely new image created from thousands of photos was not “transformed” enough from the original photo not to be considered transformative for fair use. However, a different problem arises when most or all of the images used for training the AI come from a single source. This is one of Getty’s main complaints, as they claim that Stability used millions of Getty Image photographs while also removing the captions and metadata from those images.[xii] If Stability only used Getty Images material to train their AI, then none of the produced work may be considered transformative enough for fair use. Stability potentially using the AI produced images commercially also may weigh against them in consideration of fair use.[xiii]

Four factors are given when determining whether a copyrighted work falls under fair use. These are:

(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 concerning the copyrighted work as a whole; and (4) the effect of the use upon the potential market for or value of the copyrighted work.[xiv]

These four factors are the most heavily weighed by courts in copyright infringement decisions, so most arguments should focus on these points. There are potential arguments to be made on both sides, an example being that the work created by AI is transformative, as described earlier. Other arguments that could be made are how much of the work created by AI is currently nonprofit and is for research purposes. AI is still in its infancy, meaning many of the images made by companies using AI have no intention of commercializing these images. While this will most likely change in the future, this current fact favors heavily for fair use. An argument against fair use is that the AI-generated image could potentially not exist without using the copyrighted images. Earlier, it was stated that it would be difficult to argue that a singular image significantly affects what the AI produces. Depending on the sample size used to train the AI, a particular copyrighted image could be highly important to making the AI work properly, meaning this image would be used “substantially” to create the AI-generated image. Furthermore, if AI can take copyrighted images and create very similar facsimiles under the protection of fair use, then this would undoubtedly affect the market for the original copyrighted images, weighing against fair use.

            Another issue to consider is whether the AIs themselves can own a copyright in their work. Currently, only “human authors” can hold a copyright in a product that they themselves produce or inherit.[xv] This was decided in 2018 by the 9th Circuit when there was controversy over whether a photograph that a monkey technically took could own a copyright in that photograph.[xvi] This happened when a wildlife photographer named David Slater left his camera near a reserve of wild monkeys with a mechanism for photos to be automatically taken when a button is pressed.[xvii] A monkey named Naruto pressed the button while smiling at the camera, technically making him the “author” of the photographs.[xviii] The 9th Circuit held that since all the terms used to describe an author or inheritor of a copyright “imply humanity,” animals do not fit in any of these categories and cannot own a copyright.[xix]

            Currently, there have been no other cases or statutes directly addressing this issue, leaving Naruto v. Slater as the sole authority on non-human entities and copyright ownership. This implies that AIs would be in the same category as animals as non-human entities, meaning they cannot own or hold a copyright in something. However, the court in Naruto states explicitly that it is “animals other than humans” that cannot own a copyright or sue under the Copyright Act.[xx] AI is not an animal, meaning it does not fall under this exact definition and is still in a bit of a nebulous area in terms of copyright ownership. This makes sense, as humans did not create animals, and there is, therefore, some “separation” between what an animal creates and what a human creates. The difference with AI is that humans originally made AI, and AI then, in turn, made something else based on the information humans gave it. Is this amount of separation enough to consider an AI image an “independent creation” underserving of copyright protections provided to human creators? Is copyrightable material created by an AI any different than copyrightable material made through other technologies, such as image editing software? For example, people can copyright photos they make through Photoshop, simply editing some original image with tools provided by the program to make a new or transformed image. AI is merely a program created by humans that takes thousands of images instead of a single image to create a new or transformed image. Is the fact that a program produces the image “independently” different than a human simply using a program to modify an image themselves? Or can an AI be considered independent enough to be a separate distinction from a human program? This is a much deeper issue that goes beyond the scope of copyright but is an important question that will almost certainly have many cases pertaining to it in the future.

---

[i] Jonathan Tarud, AI Model: How Does It Work?, Koombea (March 23, 2023), https://www.koombea.com/blog/ai-model/.

[ii] Id.

[iii] Eugene Dorfman, How Much Data is Required for Machine Learning?, postindustria (March 25, 2022), https://postindustria.com/how-much-data-is-required-for-machine-learning/.

[iv] Id.

[v] Getty Images (US), Inc. v. Stability AI, Inc., No. 23-CV-00135, (D. Del. May 16, 2023).

[vi] Complaint at 1, Getty Images (US), Inc. v. Stability AI, Inc., No. 23-CV-00135, (D. Del. Feb. 2, 2023).

[vii] Defendant’s Motion to Dismiss or Transfer this Action, Getty Images (US), Inc. v. Stability AI, Inc., No. 23-CV-00135, (D. Del. Feb. 2, 2023).

[viii] 17 U.S.C. § 107.

[ix] Id.

[x] 510 U.S. 569, 579 (1994).

[xi] Id.

[xii] Getty Images (US), Inc., supra note 6 at 1.

[xiii] 17 U.S.C. § 107.

[xiv] Id.

[xv] Naruto v. Slater, 888 F.3d 418, 426 (9th Cir. 2018).

[xvi] Id. at 420.

[xvii] Id.

[xviii] Id.at 424.

[xix] Id. at 426.

[xx] Id.