The use of artificial intelligence (AI) to create works of art creates interesting legal challenges to the protection of intellectual property. A copyright infringement class action lawsuit has been filed against providers of AI software that is used to generate visual works of art based on copyrighted images copied from the internet. The plaintiffs claim that defendants infringed their intellectual property by using these images to train the AI system and further allege that the images created by the AI system infringe because they “add nothing new” to the pre-existing works.

What You Need to Know:

• Generative AI is a type of AI that produces content, such as images or music, based on prompts entered by the user and the patterns and relationships the AI system recognizes in the data it was trained on.
• Stable Diffusion is a generative AI software product that is alleged to have used billions of copyrighted images taken from the internet for its training data.
• Fair use is a defense to copyright infringement. One of considerations in a fair use analysis is whether the allegedly infringing work is transformative – that is, does it add a new expression, meaning, or message not found in the original work.

​
A class action lawsuit was recently filed against Stability AI, Midjourney, and DeviantArt for copyright infringement. These companies are providers of generative artificial intelligence (AI) software. Generative AI is a type of AI that creates content, such as images, music or software code, in response to text and other prompts provided by a user. Generative AI is a type of machine learning (ML), which also is a subset of AI. How a generative AI system responds to a user’s prompts is based on how it is trained. ML systems are trained by analyzing significant amounts of data to find patterns, and generative AI programs produce an output that is based on the relationships it has “learned” about the user’s prompts and the training information.

The generative AI product at issue in the class action is Stable Diffusion, which the plaintiffs allege is maintained and sold by Stability AI. The crux of the plaintiffs’ claims is that Stability AI downloaded and used billions of copyrighted images from the internet without the artist’s authorization in order to train Stable Diffusion, and that the content generated by Stable Diffusion are derivative works that improperly compete with the training images. The plaintiffs have asserted claims against Midjourney and DeviantArt, alleging that their products rely on Stable Diffusion to generate infringing images.

In the class action, the plaintiffs assert that the images produced by the defendants’ generative AI products “add nothing new” to the training images. Whether that is legally correct remains to be seen, but it is worth noting that the images created by generative AI products are not replicas of the training images. Instead, the images that the software generates utilize a method known as “diffusion.” Diffusion is a multi-step, two-directional process in which the AI software uses the training data to generate images that resemble it, first by ruining and then rebuilding the images.

A plaintiff suing for copyright infringement has the burden of demonstrating that the defendant copied original elements of a protected work. To analyze the copying element of a claim, courts generally assess whether 1) the plaintiff has established that the defendant had access to the work at issue; and 2) the defendant’s work is “substantially similar” to the original elements found in the plaintiff’s work.

The class action plaintiffs’ contention that images generated by Stable Diffusion “add nothing new” begs the question whether the resulting content is transformative under a fair use analysis. Fair use is a defense to a copyright infringement claim that permits the unlicensed use of a copyrighted work in limited circumstances.

Section 107 of the Copyright Act provides the statutory basis for the fair use defense, and it sets out a non-exhaustive list of four factors that courts analyze whether the contested use is fair. Those factors 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 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. 17 U.S.C. § 107.

In assessing the purpose and character of the use, courts will look at the extent to which the allegedly infringing work is “transformative” – or “whether the new work merely supersedes the objects of the original creation, or instead adds something new, with a further purpose or different character, altering the first with new expression, meaning, or message.” Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 579 (1994) (internal quotations marks and citations omitted) (alterations adopted). The preamble to Section 107 of the Copyright Act identifies criticism, comment, news reporting, teaching, scholarship, or research as examples of transformative uses. 17 U.S.C. § 107. Parody is another example of a transformative use. Campbell, 510 U.S. at 580-81; Brownmark Films, LLC v. Comedy Partners, 682 F.3d 687, 693 (7th Cir. 2012).

In The Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 992 F.3d 99 (2d Cir. 2021), the Court of Appeals for the Second Circuit found that works of visual art created by Andy Warhol based on a photograph of the singer Prince were not transformative. Although this case is on appeal to the Supreme Court, it is worth discussing because the court analyzed whether visual works of art that retained essential elements of the copyrighted source images were transformative.

In particular, the Goldsmith court rejected the notion that “any secondary work that adds a new aesthetic or new expression to its source material is necessarily transformative.” Id. at 111. Instead, the court concluded that more was needed for a secondary work to be transformative, and focused on the degree to which the allegedly infringing image altered or obscured the original. In so doing, the court explained that the secondary work must embody “an entirely distinct artistic purpose” (id. at 113), and contrasted visual works that draw from numerous sources as being transformative from those which merely recast “a single work with a new aesthetic.” Id.

The class action lawsuit raises a variety of interesting issues that are likely to have a significant impact on the creation, use, and protectability of AI-generated artwork. Whether images created by Stable Diffusion are transformative is a complex question, especially considering the technical nature of the diffusion process, the large volume of data on which the AI system was trained, and because the program’s output is entirely based on and designed to resemble copyrighted images. The best advice to avoid being sued for infringement is to not utilize copyrighted works as source materials. For users of generative AI products that are trained on billions of images collected across the internet, that may not be possible.