A Registered Trademark May Help Visual Artists Protect Their Unique Styles in the Fight Against AI Text-to-Image Generators

A Registered Trademark May Help Visual Artists Protect Their Unique Styles in the Fight Against AI Text-to-Image Generators

A recent order on a motion to dismiss in the case of Sarah Andersen, et al. v. Stability AI Ltd., et al., WL 3823234 (N.D. California 2024) should raise red flags for developers of artificial intelligence (AI) text-to-image generators. Sarah Andersen involves a class of visual artists suing several owners of AI text-to-image generators under various legal theories, including trademark and copyright infringement. They claim that these owners are using their artworks and names to train AI products without their consent. In the order, the judge found that the developers of such products may be held liable for the output generated by prompts entered by end-users and for the process of creating the image-to-text generators.

Text-to-image AI generators need excessive data to create large datasets, enabling the machine learning necessary to make AI platforms effective. Proprietors of text-to-image generators, such as Stability AI, Midjourney, and DeviantArt, often deploy crawlers and bots to scrape the internet to compile these datasets, often without the permission of the owners of the copyrights and trademarks whose works are used in the process. This practice of using unapproved assets to train AI models has spawned a litany of lawsuits over the past two years, and two potentially groundbreaking findings in the recent order issued by a federal judge in Sarah Andersen could have far-reaching consequences. Refusing to dismiss many of the plaintiffs’ claims, Judge Orrick found that text-to-image output may infringe on a visual artist’s style. He also found that in the early stages of a case (depending on the facts), a defendant need not identify substantially similar aspects of its works that were infringed in order for the case to proceed.

The visual artists-plaintiffs claimed that the style of their artworks is a form of trade dress protected by federal trademark law under the Lanham Act. Their complaint noted specific features of their respective works that represent their unique styles. The plaintiffs alleged that the defendant’s use of “CLIP-guided diffusion” technology in a trade dress database, where the artists’ names are used in prompts to generate trade dress output, is likely to confuse the public with the artists’ trade dress because of their similarity. The judge agreed.

This is a seminal finding because it departs from past precedents where judges were reluctant to protect styles of artistic works under trademark law for fear they would tread on the domain of copyright law, which governs the protection of creative works fixed in a tangible medium of expression but offers no protections for styles. Moreover, the judge found that the plaintiffs’ allegations, that the defendants’ use of the plaintiffs’ names and likenesses of their works in showcasing the capabilities of the database, made a plausible claim that such conduct misleads consumers as to source and endorsement.

An essential element in proving copyright infringement is establishing “substantial similarity,” which is usually a question of fact to be decided by a jury. The jury decides whether and to what extent a qualitative and substantive amount of the original work has been appropriated by the secondary work, such that the secondary work unlawfully benefits from the author’s labors in creating the original. Such a finding by the jury means a substantial similarity exists between the protectable elements of both works.

However, in the plaintiff’s complaint, he doesn’t have to prove substantial similarity. Instead, he only has to assert a believable set of facts to establish a copyright infringement claim, i.e., that unlawful copying of protectable elements has occurred. But historically, in doing so, the plaintiff must identify those protectable elements of his work that are substantially similar to the infringing work in the complaint. That is why the judge’s order in Sarah Andersen is eye-opening. He says that is not necessarily so, given the facts of the particular case.

The plaintiffs allege that the defendants’ products use their works to train AI text-to-image generators, and this very training process incurs liability because the AI software necessarily makes copies of the artists’ artwork without their permission. As a separate basis for liability, the software owners distribute the AI models containing the artists’ works without their consent. So, even though the plaintiffs failed to identify concrete elements of their works that are infringed by the AI output, the judge said they didn’t have to at this juncture because of the unique nature of AI models, so long as a “plausible inference” could be made that infringement had transpired.

The ruling in this lawsuit is unlike other cases in the past where an alleged infringer might escape some forms of liability if it has not actively encouraged the use of its product to infringe by showing that its product is capable of “substantial noninfringing uses.” The defendants in Sarah Andersen don’t have that option even though the plaintiffs don’t dispute that the defendants’ products may be used for lawful purposes, such as generating generic artwork or styles.

Distinguishing those cases, the judge found that AI text-to-image generators are intrinsically different. They are not only developed by using the copyrighted works of others without approval, but they also generate output that includes in whole or in part protectable elements of those works when a user explicitly enters the name of the corresponding artist in the prompt field. In refusing to follow otherwise precedential decisions cited by the defendants, which Judge Orrick referred to as “‘run of the mill’ copyright cases where a showing of substantial similarity between works is required when determining whether an inference of copying can be supported,” he said those decisions “are unhelpful in this case where the copyrighted works themselves are alleged to have not only been used to train the AI models but also invoked in their operation.”

AI is poised to revolutionize every aspect of our lives, including the unauthorized use of works of visual art and source identifiers for services and goods. The Sarah Andersen case provides owners of these assets with added ammunition in their fight against developers of text-to-image generators. Visual artists may now be able to protect their recognizable styles under trademark law and, depending on the facts of their case, may not, when pleading copyright infringement, need to identify protectable elements of their works in their complaint to withstand a motion to dismiss.