Andersen v. Stability AI

The claims

Direct and induced copyright infringement over training and model distribution, Lanham Act trade dress (against Midjourney); DMCA and some state-law claims dismissed

What has happened

Three working artists filed the first class action against image-generator companies in January 2023, alleging Stable Diffusion was trained on billions of scraped images, including theirs, and that the model itself embodies infringing copies. Judge Orrick trimmed the first complaint in October 2023 but let the core theory proceed, and an amended complaint added Runway AI and more artists. On August 12, 2024, he allowed direct and induced copyright infringement claims to go forward against all four defendants, writing that the products were plausibly "built to a significant extent on copyrighted works," and let trade-dress claims against Midjourney proceed while dismissing the DMCA claims. Since then the case has been in heavy discovery: a January 2026 joint statement reported tens of thousands of pages produced by Stability and substantially complete production of Midjourney training-data materials. Plaintiffs filed a third amended complaint in early 2026, which defendants answered, and class certification remains ahead.

Key developments

• 2023-01-13 — Filed as the first proposed class action by artists against Stability AI, Midjourney and DeviantArt over image-model training.
• 2023-10-30 — Judge Orrick dismissed most claims with leave to amend, keeping the direct infringement claim against Stability; plaintiffs later added Runway AI and additional artists.
• 2024-08-12 — The court allowed direct and induced copyright infringement claims against all four defendants to proceed, plus Lanham Act trade-dress claims against Midjourney; DMCA claims were dismissed.
• 2026-01 — A joint case-management statement described extensive document production, including substantially complete production of Midjourney training-data materials.
• 2026-02 — Plaintiffs filed a third amended complaint; defendants answered in March 2026. No class certification ruling has been reported.

Why it matters for training data

Andersen established that a model trained on unlicensed images can itself be the basis of an infringement claim, not just its outputs. The induced-infringement theory reaches companies that shipped or repackaged someone else’s model, which matters to anyone reselling or fine-tuning third-party weights. Class certification, when it comes, will test whether millions of scraped works can be litigated collectively — a major swing factor in damages exposure for image-model builders and the datasets they bought.

Sources

• CourtListener docket
• Artnet on the August 2024 ruling
• Copyright Alliance case takeaways
• Joint case management statement (January 2026)
• Mesh IP Law litigation tracker

Deeper analysis

• Training-data copyright lawsuits: the landscape
• Is scraped data fair use for AI training?

← All case files Weekly lawsuit news