By QuvirAI Team — July 2026 · 9 min read

Type an artist's name into an image generator. Add "in the style of." Hit go. Ten seconds later you are holding something that looks like years of their work. They were never asked. They were never paid. That single moment is the whole fight, compressed.

One camp calls it the biggest art heist in history. The other calls it how every artist who ever lived already learned — by looking at what came before. The uncomfortable part is that the courts keep handing wins to both of them. So instead of telling you who is right, this piece lays the real evidence on the table, side by side, and lets you decide.

The case for theft: "I was never asked"

In January 2023, three artists — Sarah Andersen, Kelly McKernan, and Karla Ortiz — sued Stability AI, Midjourney, and DeviantArt. Their claim is simple to state and hard to shake: billions of images, theirs included, were scraped to train these models. No consent. No credit. No cheque.

Ortiz is not an outsider throwing stones. She designed the lead look for Doctor Strange and worked on Black Panther and Avengers: Infinity War. When she testified before the U.S. Senate Judiciary Committee, she put the grievance in words a jury could feel.

She has called it "industrial-level identity theft." Notice the grievance is about consent, not about pixels being duplicated. And the case is no longer easy to wave away. In August 2024 the judge refused to toss the core copyright claims. The lawsuit is now heading to trial, with a date set for September 8, 2026 — a courtroom test that has been three years in the making.

Step into any developer or artist forum and the mood is blunt. Here is a widely-shared comment that sums up the accusation:

The case against theft: what the courts actually ruled

Here is where the emotional version of the story runs into the legal one. When these arguments reached judges in 2025, the artists did not walk away with the clean win the headlines implied.

Start with the UK. On November 4, 2025, the High Court delivered the first British judgment on AI training and copyright in Getty Images v. Stability AI. Getty, a company with every resource to fight, actually dropped its main copyright claims during the trial. Stability won on the rest. The court accepted that Getty's images were used in training — and still found that this, by itself, was not copyright infringement. Getty walked out with a narrow trademark finding and permission to appeal. Not the verdict the "it's theft" side wanted on the record.

Then the US. In June 2025, two California rulings — Bartz v. Anthropic and Kadrey v. Meta — held that training an AI model on lawfully obtained, copyrighted work can be "spectacularly transformative" fair use. The judge's logic: learning statistical patterns to make something new is a different purpose from the original book or picture. That is a body blow to the theft argument.

There is one more piece the defense leans on, and it is older than any of these companies. Copyright has never protected style. It protects specific expression — this exact image, this exact sentence — not a look, a method, or a vibe. Painting like Van Gogh was always legal. So "make it in the style of X," on its own, sits in a zone the law has left open for a century.

The other camp answers with just as much heat. From the same kind of public thread:

The tools caught in the middle

Not every generator sits in the same seat. Some scraped the open web. One built its pitch on the opposite promise. Adobe Firefly markets itself as the "commercially safe" option, trained on licensed Adobe Stock content and public-domain work rather than a web-wide scrape.

Even that clean story picked up a stain. Reporting found that some images in Firefly's training set were themselves AI-generated, submitted by contributors to Adobe's stock library. Adobe's answer: those made up only about 5% of the data. Whatever you make of that number, it shows how hard "ethically trained" is to guarantee once the internet is your supply chain.

How artists are fighting back without a courtroom

While the lawyers argue, a research team at the University of Chicago handed artists a weapon of their own. Glaze adds a nearly invisible layer to an image that confuses models trying to learn its style. Its sibling, Nightshade, goes further — it "poisons" the data, so a model that trains on shaded images learns the wrong thing.

The adoption numbers tell you how raw the nerve is. Both tools are free. Neither will ever be sold. Artists reached for them anyway, by the millions.

The scoreboard so far

Read that scoreboard twice. The training-is-fair-use rulings were narrow and fact-specific — judges said so out loud. None of them handed AI companies a blanket permission slip, and none of them guaranteed artists a licensing fee. The one case built specifically around visual artists is the one still waiting for a jury.

FAQ