First Jury Trial Over AI Image Training Begins in California

Before a federal court in San Francisco, the trial in Andersen v. Stability AI, Midjourney and DeviantArt began on September 8, 2026. It is the first case in the United States in which the legality of training image generators on other people's works without consent will be decided by a jury, rather than settled solely through a judge's summary ruling.

From lawsuit to courtroom

Sarah Andersen, known for the webcomic Sarah's Scribbles, together with Kelly McKernan and Karla Ortiz, sued Stability AI back in January 2023, shortly after Stable Diffusion gained mass popularity among creators of computer-generated images. The artists were later joined by additional creators, and the lawsuit expanded to include more defendants, including Midjourney, DeviantArt and Runway AI. Three years of procedural disputes, motions to dismiss and discovery document exchanges culminated in the trial being set for September 8, 2026 before a jury.

The artists are represented by the Joseph Saveri Law Firm, which has for years run several parallel class-action lawsuits against AI companies, including cases involving language models. In Andersen v. Stability AI, the stakes differ from most settlements reached so far, because this is the first case in which the question of training image generators goes directly before ordinary citizens sitting on a jury, rather than ending in a settlement or a judge's own decision.

What the dispute is about

The artists claim that Stable Diffusion and similar tools were trained on billions of images scraped from the internet without the authors' consent, including their own works, and that the result is the ability to generate new images 'in the style' of a specific creator simply by typing their name into a prompt. According to the lawsuit, this is not inspiration but copyright infringement on a massive scale, because the model stores and reproduces the distinctive characteristics of other people's work.

The companies named in the suit defend themselves with a fair use argument, contending that training a model is transformative and does not directly copy any specific image, but merely learns statistical patterns from a huge dataset. In its response to the lawsuit, Midjourney explicitly called its use of training data 'the quintessence of transformative use,' also citing licenses, the de minimis principle, and the safe harbors provided under the DMCA.

Judge Orrick's earlier rulings

Before the case reached the jury, Judge William Orrick dismissed some of the claims, including those concerning violations of copyright management information under the DMCA, but allowed the main claims of direct and contributory copyright infringement and trademark infringement to proceed. In one earlier ruling, he rejected the defense's argument that generative AI operates similarly to older copying technologies such as VCRs.

This is a case where plaintiffs allege that Stable Diffusion was built in substantial part on copyrighted works - Judge William H. Orrick, Northern District of California

Why it matters for the industry

So far, no US appeals court has directly answered the question of whether training an AI model on copyrighted images falls within the bounds of fair use. AI companies have reached settlements, as Anthropic did in the books case, where the court approved a $1.5 billion settlement covering more than 482,000 works in July 2026. The Andersen trial is different, as it is the first case involving images that does not end in a settlement but instead goes before ordinary citizens weighing the facts.

The outcome of this trial could set the direction for dozens of other pending cases against companies generating images, music and video with AI. Lawyers following the case emphasize that a jury verdict, unlike a settlement, creates a factual precedent that future lawsuits and defenses in similar cases across the United States will be able to invoke.

What it means for the Polish market

For Polish creators, advertising agencies and companies using image generators, the outcome of the trial matters indirectly but tangibly. Models such as Stable Diffusion and Midjourney are widely used in Polish graphic design and marketing studios, and a ruling unfavorable to AI companies could force changes in training data licensing, which in the longer term would affect the availability and pricing of these tools in Europe as well.

In the European Union, the question of training models on copyrighted data is already partially regulated by the AI Act and copyright provisions that allow content to be reserved from text and data mining, but the American judicial precedent will still be watched by European lawyers as an interpretive signal, especially since most major model providers operate globally.

According to lawyers, the jury trial is expected to last several weeks, and a verdict could come as early as September or October 2026. Regardless of the outcome, both sides have already signaled the possibility of an appeal, meaning the final resolution of the legal question could take several more years.