Andersen v. Stability AI Ltd.

Three visual artists sued Stability AI, Midjourney, and DeviantArt in a class action alleging AI image generators were trained on billions of copyrighted artworks without permission. This is the case that will decide whether AI art generation constitutes copyright infringement, and it's moving toward discovery that will expose how AI training datasets are actually built.

Background

Sarah Andersen, Kelly McKernan, and Karla Ortiz are professional visual artists whose work was scraped from the internet and included in the LAION-5B dataset, a massive collection of image-text pairs used to train AI image generators. Stability AI used LAION-5B to train Stable Diffusion, its open-source image generation model. Midjourney and DeviantArt built products on top of similar technology.

The artists filed their class action in the Northern District of California on January 13, 2023, under Case No. 3:23-cv-00201. They alleged that the AI companies copied billions of copyrighted images to train their models, and that the resulting AI tools can generate images 'in the style of' specific artists, effectively competing with the originals.

The case was assigned to Judge William Orrick. It quickly became the leading test case for whether generative AI image tools violate artists' copyrights, drawing attention from the art community, AI industry, and intellectual property bar.

What Happened

The defendants moved to dismiss, arguing that AI training is transformative fair use and that AI-generated images are new works, not copies. Judge Orrick initially dismissed most claims in October 2023 but allowed the direct copyright infringement claim against Stability AI to proceed. The key reasoning: training the model involved making actual copies of copyrighted images, and that copying alone can constitute infringement.

The plaintiffs filed an amended complaint addressing the court's concerns. In August 2024, Judge Orrick denied Stability AI's and Midjourney's motions to dismiss additional copyright claims, significantly expanding the scope of the case. More claims survived, giving the artists a broader path to discovery.

The case is now advancing toward discovery, which will be the real battleground. The artists will get access to training data logs, internal communications about copyright concerns, and technical details about how specific copyrighted images influenced model outputs. That evidence will shape not just this case but the entire AI art copyright debate.

The Ruling

Judge Orrick's rulings addressed a core question: does copying images to train an AI model count as copyright infringement? For Stability AI, the answer was that the claim survived because training Stable Diffusion required making copies of copyrighted works. That's a straightforward act of reproduction, regardless of whether the AI outputs are themselves infringing.

The August 2024 ruling expanded the surviving claims. The court found that artists adequately alleged that Midjourney also made copies for training purposes and that both companies' outputs could constitute derivative works. This gave the plaintiffs multiple legal theories to pursue through discovery.

The court hasn't ruled on the merits yet. Fair use will be the central defense, and it requires factual development. But the fact that the case survived two rounds of motions to dismiss means the court sees viable copyright claims in AI training on copyrighted images.

Outcome: Judge Orrick initially dismissed most claims in October 2023 but allowed the direct copyright infringement claim against Stability AI to proceed. In August 2024, the court denied Stability AI and Midjourney's motions to dismiss additional copyright claims, advancing the case toward discovery.

Why This Case Matters

Andersen v. Stability AI is doing for visual art what NYT v. OpenAI is doing for text. Together, they'll define whether AI companies need permission to train on copyrighted content. The difference here is the visual medium: artists can show, side by side, their original work and AI-generated images 'in their style.' That visual comparison is powerful evidence.

The discovery phase will be groundbreaking. For the first time, a court will force AI companies to reveal exactly how their training datasets were constructed, what role specific copyrighted works played, and whether internal teams knew they were building products on unlicensed content. That factual record doesn't just matter for this case. It becomes the foundation for every AI copyright case that follows.

For the broader creative industry, this case is existential. If AI companies can train on any publicly available image without permission, artists lose control over how their work is used. If they can't, the entire generative AI art industry needs to be rebuilt on licensed datasets. There's no middle ground.

Lessons for Attorneys

Attorneys representing creative professionals should be preparing clients for the discovery outcomes. The evidence that emerges from this case will directly inform whether artists, photographers, and designers have viable claims against AI companies. Start documenting clients' copyrighted works now, including registration status, because statutory damages require timely registration.

Firms advising AI companies need to audit their training data pipelines. This case makes clear that 'we scraped it from the internet' isn't a legal defense. If the training data includes copyrighted works, the company needs a fair use argument or a license. The safest path is building toward licensed or permissioned datasets before the courts force the issue.

For IP practices generally, this case is a preview of the next decade of copyright litigation. Every industry that produces copyrighted content (publishing, music, photography, software) faces the same question. The firms that develop expertise in AI training data copyright now will own the space when these cases multiply.

The Bottom Line

Andersen v. Stability AI is the leading case testing whether AI image generation violates artists' copyrights. Key claims survived dismissal, and the discovery phase will expose how AI training datasets are built. Every attorney in IP practice needs to understand this case because it's writing the rules for AI and creative rights.