case-law

Getty Images v. Stability AI: The UK and US Rulings, Explained

By Nora Whitfield · August 8, 2026

Getty Images sued Stability AI on two continents, over two overlapping but distinct legal theories, and got two very different results. Losing the core copyright argument in one court while winning a narrower trademark claim in another sounds like a wash. It isn't. The split tells you something specific about where these cases are actually landing right now, and it isn't where most people expect.

The UK case: copyright lost, trademark won, on a technicality that matters

Getty brought its first major suit against Stability AI in the UK. The UK High Court ruled in November 2025, and Getty lost on the core copyright infringement claim, but not because the court found training on Getty's images was fair use or otherwise lawful in principle. It lost largely on territoriality: the training itself, the court found, hadn't occurred within the UK, which put it outside the reach of UK copyright law regardless of the merits of the underlying conduct.

That's an important distinction. This wasn't a ruling that AI training on licensed stock photography is legally fine. It was a ruling that this particular court, applying this particular jurisdiction's rules, couldn't reach conduct that happened somewhere else. A different training location, or a differently pled claim, could produce a completely different result even under identical facts.

Where Getty did win, narrowly, was on trademark. Early versions of Stable Diffusion, the record showed, could generate images bearing a recognizable Getty Images watermark, the exact visual mark Getty uses to identify its licensed stock photography. That's a trademark and passing-off problem distinct from the copyright question. The copyright question asks whether the underlying image was copied. This one asks whether the output falsely signals that it came from Getty, misleading anyone who sees the watermark into thinking they're looking at licensed Getty content when they aren't.

The trademark result itself was narrow. The court dismissed Getty's broader dilution claim under section 10(3) of the UK Trade Marks Act, and found only "extremely limited" infringement under the more specific sections 10(1) and 10(2), tied to the early Stable Diffusion versions that actually produced watermark artifacts in their outputs. It's a real finding of liability, the first of its kind against an AI image generator anywhere, but it's also a small one, confined to a specific historical version of the model and a specific visual artifact, not a general ruling about AI-generated imagery.

The US case: a narrower survival, but a real one

Getty's parallel US case, refiled in the Northern District of California, has followed a different track. In April 2026, the court ruled on a partial motion to dismiss: Getty's trademark, unfair competition, and trademark dilution claims survived and can proceed, while its DMCA claim, based on the removal of copyright management information, was dismissed, though with leave to amend rather than a final rejection.

Read together with the UK result, a pattern comes into focus. The straightforward "you copied my copyrighted images to train your model" theory has struggled in both courts, for different procedural reasons in each. What's actually surviving and moving forward is the trademark and provenance-adjacent theory: did the output falsely represent its source, did it carry someone else's mark, did it create confusion about where the image actually came from.

Judge Trina Thompson's April 2026 ruling on the DMCA claim is worth understanding on its own, because the reason it failed is instructive. The Digital Millennium Copyright Act's false copyright-management-information provision doesn't just require showing that identifying information was altered or stripped. It requires showing the defendant acted with a specific intent, what the law calls scienter, roughly meaning the defendant knew or had reason to know the alteration would conceal infringement. Getty's complaint, as filed, didn't plead facts specific enough to meet that intent standard, so the claim was dismissed, though without prejudice, meaning Getty gets a chance to refile with more detail. That's a pleading problem, not necessarily a merits problem, and it's a common failure point in DMCA claims against AI companies generally: proving that a specific technical outcome (a watermark surviving into an output, or metadata disappearing during training) was the product of intentional concealment, rather than an unintended side effect of how a model processes its training data, is a genuinely hard evidentiary bar to clear at the pleading stage.

Why the watermark detail matters more than it sounds like it should

It's tempting to read the watermark issue as a minor technical footnote next to the bigger copyright fight. It isn't. It's arguably the more durable legal theory to come out of this entire litigation so far. A copyright claim over training data forces a court into the hardest, most unsettled question in this entire area of law: is training itself a fair use. A trademark claim over a reproduced watermark sidesteps that fight almost entirely. It's a much older, much more settled body of law, and courts in two different countries have now shown more willingness to let that theory proceed than the underlying copyright theory.

What this means for AI authors

If you create or license visual work, this case is a useful signal about where the actual legal traction is right now: not in "AI trained on my images without permission," which remains a genuinely hard fight to win outright, but in provenance and source-identification claims, whether an AI tool's output falsely implies it came from you, carries your mark, or misrepresents its origin. That's a meaningfully different, and in some ways more winnable, angle than a pure training-data copyright claim.

This is also directly relevant to how you should think about your own authenticity claims. If you're producing work and want to be able to show, clearly and verifiably, that a given piece actually came from you, provenance and content-authenticity tools (things like Content Credentials and C2PA, covered elsewhere on this site) are doing exactly the kind of work that turned out to matter in this case: not proving what was copied, but proving what's genuinely yours and where it actually came from. The Getty case is a real-world example of how much that distinction can matter once a dispute actually reaches a courtroom.