Copyright in the Age of Generative AI: Legal Gaps and Artist Vulnerability
It started with crude parodies on Twitter and Reddit. Cartoonist and illustrator Sarah Andersen’s signature comics featuring a protagonist—a wide-eyed girl with black bangs wearing a striped shirt, a nod to her favorite character in “Calvin and Hobbes”—were first vandalized by alt-right groups who twisted her relatable humor into malicious content. They inserted violent texts advocating genocide and racism into her work without her consent to distribute on social media platforms. Andersen’s creative platform, as well as those of similarly impacted digital artists, was tainted by distorted texts and streaks.
The launch of image-generating software such as Stable Diffusion by Stability AI heralded a new era of manipulation and theft of artistry. When fed Andersen’s name in a prompt, Stable Diffusion recalled unmistakable patterns in her drawing style, such as the wide eyes and striped shirt of her signature character. While Stable Diffusion did not produce exact copies of Anderson’s original works, it exhibited potential to mimic her style convincingly. In a New York Times op-ed, she expressed, “I felt violated. The way I draw is the complex culmination of my education, the comics I devoured as a child, and the many small choices that make up the sum of my life.” Stable Diffusion threatened artists’ livelihoods by copying personal details in their artistic styles at the touch of a button. Driven to protect her intellectual property rights and creative authority, Andersen joined forces with other artists to challenge the use of unconsented, uncompensated art in the development and operation of image-generating software.
On January 13th, 2023, Andersen, along with two visual artists in the Northern District of California, filed a class-action lawsuit against Generative Artificial Intelligence (gen AI) platforms. Gen AI is broadly defined as a machine-learning model that creates synthetic content based on training data such as text, images, video, and audio. The plaintiffs’ claims of copyright infringement, Digital Millennium Copyright Act (DMCA) violations, and false endorsement were based on the use of their works in training datasets for Stability AI’s Stable Diffusion and DreamStudio, DeviantArt’s DreamUp, and Midjourney Inc.’s own gen AI tool. As the first lawsuit where artists rallied together to challenge copyright infringement against gen AI companies, Andersen v. Stability AI Ltd. sets a critical groundwork for cases on intellectual property rights. Many authors and visual artists have since taken similar steps by filing class action lawsuits.
On October 30th, 2023, after a series of motions and oppositions, the district court largely dismissed the artists’ lawsuit with leave to amend (permission to modify the initial filing), except for Andersen’s claim of direct copyright infringement against Stability AI. Only Andersen’s direct infringement claims survived because she had registered some of her works with the U.S. Copyright Office, a requirement for initiating an infringement action in court. The artists, except Andersen, were also unable to provide sufficient evidence of identical output images that would support a claim of removed copyright management information (CMI). To show a violation of the DMCA 1202b, which states “No person shall, without the authority of the copyright owner or the law—(1) intentionally remove or alter any CMI, (2) distribute or import for distribution CMI… or (3) distribute…copies of works…knowing that CMI has been removed or altered,” the artists needed to prove that gen AI software intentionally removed CMI, or distributed works knowing that CMI had been removed. In other words, the plaintiffs had to prove that the AI had deliberately stripped CMI from their original works, or that the developers built the AI with the intent to facilitate copyright infringement. The defendants countered that the AI was designed to generate new, distinct images—not to replicate those in the training data. The court agreed with the defendants, ruling that the AI-generated images did not infringe on copyright as they were not intended to be substantially similar to the artists’ protected works.
Proving intent of infringement is notoriously difficult in copyright cases, even without the complexity of gen AI. In Philpot v. WOS, Inc., freelance photographer Larry Philpot alleged that media company WOS removed his CMI by scraping the metadata off his photos in violation of the DMCA 1202b. In their article publications, WOS had removed metadata from two of Philpot’s photographs and attributed them to “Wikimedia Commons” and Philpot’s Wikimedia username. In response, WOS’s CEO O’Dwyer testified that WOS “did not know that the photos contained metadata when it pulled the files off Wikimedia” and argued that Philpot failed to present evidence concerning WOS’s malicious intent or actual knowledge. The District Court for the Western District of Texas agreed, finding that Philpot did not meet his burden to prove that WOS intentionally removed his CMI from the photos.
In a rare case where the plaintiff successfully showed intent of infringement in gen AI, during the ongoing trial of Getty Images v. Stability AI, Getty Images filed a complaint based on evidence of Stability AI’s intent to remove or alter CMI. Getty Images showed images generated by Stability AI that, while not identical, contained infringed versions of Getty Images’ watermark. In some generated images, modified versions of Getty Images’ watermark appeared on “bizarre or grotesque synthetic imagery that tarnishes Getty Images’ hard-earned reputation.” Having a recognizable trademark allowed Getty Images to allege claims on DMCA 1202a, 1202b, and direct copyright infringement with incontrovertible evidence. However, for most independent concept artists publishing digitally without a trademark watermark, it would be difficult to allege copyright infringement through the DMCA 1202b without providing smoking gun evidence of CMI removal or alteration. Such was the case for the visual artists in Andersen v Stability AI Ltd.
In Andersen’s case, out of all her web-scraped cartoons, only her registered works could be used to proceed with the class action. For artists, applying for registration with the U.S. Copyright Office requires an application, a filing fee, and a copy of the work. The costly and long process severely disadvantages digital artists from securing intellectual property rights against gen AI companies. This undertaking poses a financial barrier to cartoonists, graphic illustrators, and photographers who make a living from frequently published concept art. Moreover, artists have slim chances of proving intent of infringement without heavily branding their own artwork in the likes of Getty Images.
The court’s dismissal of the plaintiffs’ copyright claims shifts the burden of protecting intellectual property rights into artists’ hands instead of holding corporations accountable for using unconsented artwork to train gen AI models. Andersen v Stability AI Ltd. and other ongoing lawsuits against gen AI indicate two urgent needs: a broader reformation in regulating data in open-source training sets and a radical revision in compensating artists in a recast copyright business model. Emerging legislation must include expert examination of how copyright law applies to machine-learning code and training datasets, as well as progressive imagination on how fair compensation forms for artists might evolve. Beyond three visual artists, Andersen v. Stability AI Ltd. paints the outlook for future intellectual property rulings in the face of technological development.
Edited by Ananya Bhatia and Leah Druch