Andersen v. Stability AI

Andersen v. Stability AI · editorial illustration

Background

The original class-action against generative image models. The case that prompted the creation of this website in 2023. A 2024 ruling allowed direct infringement and DMCA claims to proceed; induced infringement and right-of-publicity claims were substantially narrowed.

“Stable Diffusion is merely a complex collage tool.”

— Original complaint, 3:23-cv-00201, ¶ 11 (since narrowed)

The Claims

1. Direct Copyright Infringement. Claim that ingestion of LAION-5B and related datasets into model training constitutes unauthorized reproduction of copyrighted artworks.
2. Vicarious Infringement. Theory that model operators are liable when users generate outputs substantially similar to plaintiffs' work.
3. DMCA § 1202 Violations. Removal of Copyright Management Information from training data.
4. Right of Publicity. State-law claims that generating works "in the style of" a named artist misappropriates their persona.

The Defense

Defendants argued that training is transformative use, that diffusion models do not "store" training images but learn statistical relationships, and that style is not copyrightable under long-standing doctrine.

Analysis

Judge Orrick's multi-round motion practice significantly narrowed the case. The "collage tool" framing was rejected; many plaintiffs were dismissed for failure to register copyrights before filing.

What survived matters: direct infringement claims for plaintiffs with registered works and DMCA § 1202 claims are moving to discovery. Vicarious claims and right-of-publicity theories based on style alone were largely dismissed.

The case is unlikely to answer whether training is fair use — it will answer whether plaintiffs can prove their works ended up in the training set and whether CMI was stripped.

Timeline