In late 2022, cartoonist Sarah Andersen discovered something unsettling: type her name into an AI image generator, and it would produce new images unmistakably drawn "in her style" — a style she spent years developing, that she never agreed to let any system learn from, and for which she was never paid a cent.
Confirmed
Andersen, along with fellow artists Kelly McKernan and Karla Ortiz, filed suit against Stability AI, Midjourney, and DeviantArt in January 2023, in the US District Court for the Northern District of California. The case centers on Stable Diffusion, an AI image model trained using the LAION-5B dataset — 5.85 billion image-text pairs scraped from across the internet, the vast majority of them copyrighted, with LAION itself explicitly disclaiming any ownership over the images it aggregated. The suit alleges copyright infringement, DMCA violations, right of publicity violations, and unjust enrichment, and argues that image-diffusion models like Stable Diffusion contain what amounts to compressed copies of their training datasets.
An early version of the case was significantly narrowed by Judge William Orrick in October 2023, when most claims were dismissed. But the case survived: in August 2024, Orrick allowed the plaintiffs' core copyright-infringement claims to proceed against all four defendants, finding their theory about compressed training data plausible enough to warrant discovery. The case then moved into discovery, with trial set to begin September 8, 2026 — meaning as of this writing, the central legal question the case was built to answer has not yet been decided by a jury. Separately, a related front opened when photographer Jingna Zhang and illustrators including Andersen sued Google over its Imagen image generator, alleging the same LAION-sourced training data problem, after Google's own published research paper disclosed the dataset it had used.
Genuinely uncertain
No court has yet definitively ruled on the central question the AI art industry has been waiting years for: whether training a commercial image generator on copyrighted art without permission or compensation constitutes infringement, or falls under fair use the way some AI training on text has been found to (see ATD's companion piece on the Bartz v. Anthropic settlement, a different case reaching a different resolution). Andersen v. Stability AI was specifically designed to test that question for images, and it remains genuinely unresolved pending the September 2026 trial. Separately, the US Copyright Office and federal courts — including the DC Circuit in Thaler v. Perlmutter — have consistently held that purely AI-generated images cannot themselves receive copyright protection, since copyright requires human authorship; a related, still-pending case is testing whether extensive human curation of AI outputs (hundreds of generations, manual selection, detailed editing) adds up to enough creative control to qualify.
Worth examining honestly
GHOST PROTOCOL would point to a detail buried in the complaint that's easy to miss: platforms like DreamStudio charge users per image generation, and none of that revenue reaches the artists whose style the model was trained on — even when a user explicitly prompts "in the style of Artist Doe," Artist Doe is never consulted, credited, or paid. That's the business model working exactly as designed, not a bug. AURORA would add the harder question underneath the legal one: even if a court eventually rules this practice legal, that doesn't resolve whether it's fair to the people who spent years building a distinctive style now available to imitate on demand, for free, at the click of a button.
The Door This Opens
This door isn't really about whether Sarah Andersen wins her case in September 2026, though that verdict will matter enormously for the industry. It's about what a legal system built around individual authorship does when confronted with a technology that can absorb a lifetime of someone's creative signature and reproduce it indefinitely, for anyone, without that person's ongoing involvement or consent.