Andersen v. Stability AI

What the artists are claiming

In January 2023, three working illustrators — Sarah Andersen, Kelly McKernan and Karla Ortiz — filed the first class action brought by visual artists against image-generator companies. Their theory reaches further than the usual complaint that a specific output looks too much like a specific artwork: they argue that Stable Diffusion itself, as software, is built from billions of scraped images, theirs included, and that training the model was an act of copying regardless of what any single output looks like.

A copyright claim requires ownership of a registered work and unauthorized copying of protected expression. The artists say the copying happened when their images were downloaded, processed and used to adjust the model’s weights — a technical process, but one they argue counts as reproduction under the Copyright Act. Stability has argued that training extracts statistical patterns, not retained expressive copies.

How the case survived dismissal

Judge William Orrick trimmed the first complaint in October 2023, dismissing several claims but preserving a direct-infringement theory against Stability. Plaintiffs amended, adding Runway AI and more artists. On August 12, 2024, Orrick allowed direct and induced copyright infringement claims to proceed against all four defendants, finding the products were plausibly “built to a significant extent on copyrighted works.” That is a motion-to-dismiss ruling, not a finding of liability — it means the claims were plausible enough to reach discovery, not that infringement has been proven.

The induced-infringement theory is the more novel piece: it targets companies, like Midjourney and DeviantArt, that did not train Stable Diffusion themselves but distributed or built products on top of it. If that theory holds up, liability could reach any company that ships or fine-tunes someone else’s model without knowing, or willfully ignoring, what it was trained on.

“the products were plausibly “built to a significant extent on copyrighted works””— Order on motion to dismiss, N.D. Cal., Aug. 12, 2024

Where discovery stands

The case has been in heavy fact discovery through 2026. A January 2026 joint case-management statement described tens of thousands of pages produced by Stability and reported that Midjourney’s production of training-data materials was substantially complete. Plaintiffs filed a third amended complaint in February 2026, which defendants answered the following month. Class certification — the fight over whether millions of scraped images can be litigated as one case rather than artist by artist — has not yet been decided.

The trade-dress wrinkle

Alongside the copyright counts, Midjourney faces a Lanham Act trade-dress claim, alleging its outputs can mimic a specific artist’s recognizable visual style closely enough to confuse the market about who made the work. Orrick let that claim proceed in the same August 2024 order while dismissing the plaintiffs’ DMCA claims outright.

Why it matters for training-data licensing

Andersen is the case most directly testing whether a trained model can itself be the infringing article, not just its outputs — a theory that, if it survives summary judgment, would pressure anyone holding or reselling model weights trained on unlicensed images. For data buyers, the induced-infringement theory is the sharper edge: liability does not stop at the company that scraped the images, it can reach downstream distributors, which is exactly the chain-of-custody problem licensing is supposed to solve.