Judge trims artists' copyright lawsuit against Midjourney and Stability AI

Generative AI just survived its first major legal stress test in federal court, but the underlying threat to model training is far from gone.

California Federal District Judge William Orrick issued the first ruling on a motion to dismiss in the high-profile copyright lawsuit Andersen v. Stability AI. The suit, filed by visual artists Sarah Andersen, Kelly McKernan, and Karla Ortiz, targeted Stability AI, Midjourney, and DeviantArt for scraping billions of web images to train image generators.

The judge trimmed away most of the artists' claims, completely dismissing Midjourney and DeviantArt from the case for now. But he kept the biggest grenade on the table: Stability AI must still face direct copyright infringement claims for scraping copyrighted work to train Stable Diffusion.

Courts reject the "AI outputs are compressed copies" theory

The plaintiffs tried to argue an expansive theory: that Stable Diffusion stores compressed copies of billions of training images, making the software itself an illegal derivative work and turning every single image it generates into automatic infringement.

Judge Orrick was openly skeptical. He noted that the artists themselves conceded AI outputs rarely match any specific training image closely. In copyright law, you generally cannot claim infringement without showing substantial similarity.

Orrick pointed out that the plaintiffs failed to explain technically how Stable Diffusion operates, demanding they clarify whether the model actually stores compressed files like JPEGs or merely learns mathematical weights and parameters. Claims against Midjourney and DeviantArt were dismissed with leave to amend, while McKernan and Ortiz were dismissed entirely because they had not registered their works with the U.S. Copyright Office.

Stability AI still has to answer for the training data

While downstream app makers dodged a bullet, Stability AI is not off the hook. Judge Orrick ruled it is simply too early in the litigation to decide whether copying happened during model training.

The allegation that Stability scraped and downloaded copyrighted images without permission to build Stable Diffusion is enough for the direct infringement claim to proceed. Stability claims training does not store images, but merely calculates visual attributes like lines, colors, and shades.

The artists' legal team, led by Joseph Saveri and Matthew Butterick, celebrated keeping their core claim alive and promised to refile next month with deeper technical evidence.

Why it matters

We think this ruling brings much-needed technical realism to the legal debate around AI.

For artists and creators, this is a clear reality check. Sweeping moral arguments that every AI output is stolen work will not fly in court without proving substantial similarity. Suing AI companies requires technical precision and basic administrative homework, starting with formal copyright registration.

For AI startups, downstream builders just got breathing room. By refusing to hold DeviantArt or Midjourney automatically liable simply for running Stable Diffusion, the court signaled that bundling an open-source model does not instantly make you guilty of wholesale infringement.

However, the existential risk for foundation model developers remains wide open. The court preserved the core question: does scraping copyrighted images to build training sets constitute infringement, or is it protected under fair use? The survival of that single claim means the legal foundation of generative AI still rests on shaky ground.