Artificial Intelligence is coming to art near you! And with AI’s entrance into the field through systems like Stable Diffusion, DALL-E, and Midjourney, copyright infringement lawsuits are the next wave as we try to sort out the technology and the rights of persons affected by the technology.
The use of AI in generating art online has become increasingly popular. Many artists and designers are using AI tools to create new and unique pieces of art, from digital paintings to animations and more. The problem is how these AI systems learn how to respond to text prompts from system users.
One approach to training AI systems is to use images scraped from the internet – regardless of the copyright status of the image – and to use this library of images to inform the algorithm on how to interpret a text prompt from a system user. So, to simplify, if a system user enters a text message such as “please make an image of a cow standing in a green field on a rolling hill,” the AI algorithm would need to have reference material of what a cow, green field, and rolling hill looks like. The reference images that are indexed in relation to these words and phrases help the AI to learn how to respond to the text prompt and generate a new image.
The problem comes in with the images that are used by the AI in its original training, as some aspect of these images must be re-used by the AI to respond to a user text prompt. If the copyright owner of the source image has not dedicated the image to the public domain, or otherwise licensed the work openly, an AI that works based on combinations of these protected works may infringe a copyright right (such as the exclusive right of duplication, derivative work, etc.) under section 106 of the Copyright Act.
And several lawsuits along these lines have been filed, challenging the use by AI companies of these original source images. For example, Anderson v. Stability AI is a putative class action lawsuit brought by artists alleging infringement against Stability and several other AI image generation businesses. Getty Images has also filed a similar type of infringement suit. In Getty’s case, the images it owns have a proprietary watermark that permits the company to track unlicensed use of works, and alleges that “millions” of its images have been used by companies like Stability AI without license or consideration.
This leads to the unanswered question of whether the use of these protected works is infringing, and if it is, whether the use could be a defensible fair use under section 107. A key question for the source works harvested from internet sources is whether they are sufficiently original to be protected. The bar of originality through court decisions here is relatively low. Feist v. Rural Tele., 499 U.S. 340 (1991). Absent slavish copying of an original, most photographs will have some originality, at least as to photographer choices about composition and lighting, though the copyright interest in photographs of common objects and scenery must be relatively narrow, which is another way of saying that an infringing work must be nearly identical for liability to accrue. However, there really is no fixed formula of originality of all photographs and the degree of similarity for a subsequent work to be infringing. Such issues require a judgment on a case-by-case basis. For example, are Warhol prints based on a photograph of the late musician, Prince, sufficiently different to not be a copy of the protectible aspects of the photograph? The Second Circuit judged them to not be sufficiently different and to also not be sufficiently transformative to be protected under fair use. Even a casual reading of the literature on fair use would find the wide range of decisions and the absence of simple rules in this area. Andy Warhol Foundation for the Visual Arts v. Goldsmith, 992 F.3d 99 (2d Cir. 2021) (oral arguments heard at US Supreme Court Oct. 12, 2022).
Another problem of duplication involves the doctrine of scenes a faire, which essentially is the idea that genres or subject matters of works will necessarily share some common ground. Feist, 499 U.S. at 340. All works of penguins will share commonality in the “penguinness” of the birds pictured, therefore my photograph of penguins cannot prevent all others from taking their own photo of that bird at the zoo or in the wild. Teaching AI what image relates to the word “penguin” necessarily means showing the AI many penguin images for it to construct some generalized model of “penguin.” Therefore, copyright holders would not be able to prevent AI from generating images of penguins, even images that are based on images made by others, as all pictures of penguins necessarily include aspects of the bird’s form, markings, and size that make it a penguin rather than other birds or animals.
Another unresolved issue with this situation is the idea of sampling. The Ninth Circuit discussed several sampling cases – where a small portion of a work is used by another such that the average observer would not notice the appropriation – and explained that such sampling is generally not actionable infringement. Bell v. Wilmott Storage, 12 F.4th 1065 (9th Cir. 2021) (discussing Newton v. Diamond, 388 F.3d 1189 (9th Cir. 2004). A court could view the AI use of copyright materials of other as a mere sampling of that protected material, and therefore not actionable if the AI were to stitch together hundreds of pieces of digital files to render a response to a user-supplied text prompt.
A final fair use issue has to do with Google’s admitted copying of millions of paper books and indexing of those books into Google’s index of these works, without permission or license from the copyright holders. The Second Circuit ultimately affirmed the finding of fair use in this case, partly on the grounds that works that were still within copyright, while indexed in the database, only appear to searchers in “snippet” form that displays a limited portion of the relevant book page with the search result. Author’s Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015). In addition, many courts will examine whether the use of the plaintiff’s work is “transformative,” though this determination often can be viewed as subjective. For example, the second circuit did not find Andy Warhol’s use of Goldsmith’s photograph of Prince as transformative, but the same circuit has found numerous other appropriationist artist’s use of another’s work as transformative, such as in Cariou v. Prince, but not in Rogers v. Koons, yet confoundingly Koons’ use of a different photograph in Blanch v. Koons was sufficiently transformative. Andy Warhol, supra; Cariou v. Prince, 714 F.3d 694 (2d Cir. 2013); Rogers v. Koons, 960 F.2d 301 (1992); Blanch v. Koons, 467 F.3d 244 (2d Cir. 2006). Perhaps a court could view AI’s use of photographs scraped from the internet as sufficiently transformative, as AI generates a “new” image based on the user text prompt from relevant reference material.
While the source materials are photographs that are published on the internet, a court would most likely view them as within the core protection of the Copyright Act, so this factor would likely favor photographers whose works have been used by AI to generate images. Cariou, 714 F.3d at 710.
The third factor, the amount and substantiality of the protected work used by AI to generate the subsequent image to the AI user, is less certain as to which party is favored. Id. We do not really know how much of the scraped images the AI actually uses in the finished product it displays to users of these various systems, or whether the AI has used more than is necessary to accomplish its transformative purpose.
Finally, courts would have to examine whether the AI’s use “takes away from”usurps” the existing market for the original photographs. Id. at 709. The market for AI-generated art is relatively new. DALL-E does have a paid subscription level for image generation beyond a minimum number of free images; whereas Midjourney is a subscription-based system that starts at $96 per year. What remains to be determined is whether these paid systems reduce the market value of works in Getty Images’ database. For example, are AI users merely generating images with the AI system that they would otherwise just license from Getty Images? While this may seem unlikely from a cursory use of the AI systems, what evidence offered in court of market impact may support either party’s contentions here. However, if the AI systems are generating income, the door is open for these systems to enter into some licensing system with the works they have used to go “legitimate.”
Stay tuned for what’s next and further market disruption by AI systems available and yet to come to market.