US judge lets artists' claims against image generators proceed

Background

Text-to-image systems generate pictures from written prompts. Building such a system usually involves large collections of images, and whether using other people's pictures that way is lawful has been disputed in US courts.

The lawsuit was brought by three illustrators — Sarah Andersen, Kelly McKernan and Karla Ortiz — who initially sued Stability AI, Midjourney, DeviantArt and Runway AI in January 2023^[1]. Seven other artists joined an amended complaint in November 2023^[1]. The case is Andersen v. Stability AI, in the US District Court for the Northern District of California^[1].

What the judge decided

In August 2024, US District Judge William Orrick ruled that the artists could continue to pursue some of their claims that the four companies' image generation systems infringe their copyrights^[1]. The judge said the artists plausibly argued that the companies violate their rights by illegally storing their works on their systems^[1].

The outcome was mixed rather than a clean win for either side. Orrick refused to dismiss related trademark-law claims, but he threw out claims of unjust enrichment, breach of contract and breach of a separate US copyright law^[1]. In other words, part of the case was allowed to go forward and part of it was cut away.

What was left open

Two central questions were not resolved by the decision. It did not address the artists' core claim that misuse of their work to train AI systems directly infringes copyright, and it did not address the companies' fair use defence^[1]. Those questions therefore remained outside the ruling.

One strand of the decision was more specific. Orrick wrote that the plausible inferences are that Stable Diffusion creates copyright infringement when operated by end users and was created to facilitate that infringement by design^[1]. That reasoning is about how a model behaves when people use it, as well as about the purpose behind how it was made — a narrower idea than the broader argument that training on the works is itself infringing.

Why it matters

The ruling did not end the litigation, and it did not settle whether training image generators on copyrighted pictures is lawful. What it did was keep a portion of the artists' case alive past the point where it could have been dismissed outright, and let trademark claims continue alongside it^[1].

For readers following the wider debate over generative models, the decision is a useful marker of where the argument stood. Courts were willing to let claims about how works are stored on company systems, and about what a model was designed to do, move ahead, while the biggest question — training and fair use — was still unanswered. The case therefore continued in the same court, with those issues still to be argued.

The dispute also shows how copyright cases against AI companies tend to proceed in pieces: a single complaint can carry several legal theories, and a judge may accept some while rejecting others. The August 2024 ruling sorted the theories in this case into those that could proceed and those that could not^[1], leaving the rest for later stages.