AI Art Training Goes on Trial, and a Jury Will Decide This Time

For three years, the fight over whether AI companies can train image models on artists' work has mostly ended in settlements or stayed stuck in procedural arguments. September changed that in the United States, in two different courtroom settings, and the outcomes could shape how every generative image, video, and music tool is built from here.

The First Jury Trial on Image Training

On September 8, a trial began in federal court in San Francisco in Andersen v. Stability AI, Midjourney, and DeviantArt. It is the first case in the United States where a jury, rather than a judge ruling on a legal motion, will decide whether training image generators on other people's work without consent violates copyright law.

The case started in January 2023, when illustrator Sarah Andersen, known for the webcomic Sarah's Scribbles, and artists Kelly McKernan and Karla Ortiz sued Stability AI shortly after Stable Diffusion became popular among creators. More plaintiffs and more defendants joined over time, including Midjourney, DeviantArt, and Runway AI. Three years of motions, dismissals, and discovery led to this trial.

The artists argue that Stable Diffusion and similar tools were trained on billions of images scraped from the internet without consent, including their own, and that the result is a system that can generate new images "in the style" of a named creator from a prompt. They call this mass infringement rather than inspiration, on the theory that the model stores and reproduces the distinctive features of the works it learned from. The companies answer with fair use, arguing that training is transformative and does not copy any specific image, and Midjourney described its use of training data as the essence of transformative use.

Judge William Orrick, who handled earlier stages, dismissed some claims, including ones about removed copyright metadata under the DMCA, but let the core direct and contributory infringement claims go forward. He rejected the defense's comparison of generative AI to older copying technologies like the VCR. That is worth noting: the judge who narrowed the case still let the central question reach a jury.

Why a Jury Changes Things

A settlement creates money changing hands and a boilerplate statement. A jury verdict creates a factual precedent that later cases can point to. That is the difference the artists' lawyers are counting on, and it is why this trial matters beyond the parties involved.

The backdrop is a string of settlements. Anthropic agreed to a 1.5 billion dollar settlement covering more than 482,000 works, approved by a court in July 2026. Settlements resolve a case without answering the hard legal question. If the Andersen jury returns a verdict, that question gets an answer, at least for the Northern District of California, and dozens of pending cases against image, music, and video companies will read it as a signal.

The Same Question in New York

On the other side of the country, a judge in Manhattan is weighing the same issue for text. In the consolidated litigation known as In re OpenAI, the New York Times, a group of authors including John Grisham, George R. R. Martin, Jodi Picoult, and others, and OpenAI with Microsoft all filed for summary judgment in September, asking the judge to decide the fair-use question on the record instead of sending it to trial.

The authors argue that their books were copied without permission, including downloads from a shadow library, and that generative AI competes directly with human-created work by producing new material that can substitute for it. They warn that AI is diluting the market for books. OpenAI and Microsoft counter that training extracts broad statistical patterns rather than protected expression, and that the record shows no substitution for the actual books. Microsoft framed it as the difference between the speculation that started the case and what discovery later showed.

Judge Sidney Stein must now choose between two earlier, conflicting approaches from San Francisco. One judge called Anthropic's use of books "quintessentially transformative." Two days later, another ruled for Meta but warned that training would not be fair use "in many circumstances," raising the concern that generative AI could flood the market with content competing against human creators. The Supreme Court has not touched the question, and a clear appellate answer could be years away.

What It Means for Builders and Creators

For the companies shipping generative models, the risk is now concrete rather than theoretical. A ruling against fair use would force changes to training data, licensing deals, or model design, and those costs flow into the price and availability of tools. The EU already regulates this area through its AI Act and copyright provisions that let creators reserve their work from text and data mining, so a US precedent would be read there as well.

For creators, the two tracks point in opposite directions. A jury victory would validate the claim that training on an artist's work without consent is infringement. A fair-use win for the model makers would cement scraping as legal, at least under current US law. Either way, the era of quietly training on whatever is online is under real pressure, and the pressure is coming from ordinary citizens sitting in a jury box rather than regulators or executives.

The Stakes Beyond the Courtroom

The reach of these cases extends past the parties. For creators in markets far from either courtroom, the outcome matters indirectly but concretely. Models like Stable Diffusion and Midjourney are widely used in design and marketing studios around the world, and a ruling that forced changes to training-data licensing would eventually affect the availability and pricing of those tools everywhere, not just in the United States. The European Union already partially regulates the question through its AI Act and copyright provisions that let rights holders reserve their content from text and data mining, and lawyers there are watching the US cases for guidance, since most major model providers operate globally.

There is also a business signal in the way the cases have unfolded. Anthropic's 1.5 billion dollar settlement, covering more than 482,000 works, set a benchmark for what training on copyrighted material might cost when it is not defended as fair use. That number inevitably shapes how the next round of licensing deals gets priced, and whether companies decide it is cheaper to license data up front or to litigate later. The Andersen jury, if it reaches a verdict, adds a factual data point to that calculation in a way a settlement never does.

The honest expectation is still a messy outcome. Three years in, the field has produced contradictory rulings, partial dismissals, and settlements that resolve money without resolving law. A jury verdict or a summary judgment will add one more piece rather than settle the whole board. What has changed in September is that the questions are finally being answered on the record, in a courtroom, with consequences. For an industry that has spent three years building on legal assumptions it could not confirm, that is a threshold, and there is no going back across it.