Reading time: four minutes

The long-awaited Getty Images v. Stability AI ruling finally landed, and wow – what a plot twist. After months of breathless speculation that this case might become the grand legal showdown that defined how UK copyright law applies to AI training, the judgment ended up looking very different.

The big punchline: “there are no copies in the model.”

At the heart of Getty’s secondary copyright claim was the idea that Stable Diffusion itself, the trained model, was basically a giant machine stuffed full of stolen images. Getty argued that because its photos were used during training, the resulting model must contain or embed reproductions of its works. That would make the model an “infringing copy” under the Copyright, Designs and Patents Act, meaning Stability AI would be liable simply for bringing the model into the UK.

The court shut that down completely. With striking clarity, it said the Stable Diffusion models don’t contain Getty’s images – not as files, compressed versions or temporary fragments. The model weights are simply the mathematical patterns the system learned during training. They’re not pictures, they’re not condensed JPEGs and they’re not reproductions of copyrighted works in any normal or legal sense.

This finding obliterated Getty’s secondary infringement theory. If there are no copies in the model, then the model can’t be an infringing copy. And if it isn’t an infringing copy, importing or possessing it in the UK doesn’t trigger secondary liability. Getty’s most important live claim simply evaporated.

This one factual conclusion alone makes it extraordinarily difficult for anyone to argue in future cases that a diffusion model “contains” copyrighted works. And because it’s a factual finding, not a legal reinterpretation, it’ll be very hard to overturn on appeal.

A twist: the court says an ‘article’ can be intangible

One unexpected twist in the judgment has set off alarm bells in legal circles. The court decided that an ‘article’ under UK copyright law doesn’t have to be physical. It can be intangible, a digital file, a cloud-based system or essentially any non-physical medium in which a copy of a work could exist.

On the surface, that sounds like a blow against AI developers. In reality, it didn’t help Getty at all, because the court paired this idea with the ‘no copies in the model’ finding. But the idea itself, that an ‘article’ can be intangible, raises major questions.

For nearly 50 years, UK copyright law was aligned with EU copyright law. Under the EU’s Information Society Directive and the Court of Justice’s case law, the right of distribution applies only to tangible objects. Downloads, streams, cloud access, transmissions, these all fall into completely different legal categories. EU law drew a bright like: goods are tangible, services aren’t.

By declaring that an article can be intangible, the UK High Court appears to have wandered away from that EU framework without acknowledging the conflict, which it arguably doesn’t have the authority to do. Only higher courts can override retained EU law.

This little interpretive detour is almost certainly going to be interrogated in future cases, if not appealed directly.

Getty’s case narrowed to almost nothing

What makes this judgment even more surreal is that Getty itself had already abandoned the most important parts of its lawsuit. In the middle of the proceedings, Getty accepted that the model wasn’t trained in the UK and that it couldn’t produce a meaningful sample of infringing outputs. Once that happened, the primary copyright and database-right claims collapsed.

The court was left to decide only the secondary infringement question and a handful of trademark accusations. The big existential questions, whether scraping the web for training data infringes UK copyright, whether training itself is reproduction and whether outputs can infringe, were all left floating unanswered.

So, the very dispute the world was watching for… never actually happened.

Hosted access vs downloads: the legal distinction that really matters

One of the most practically important elements of the decision concerns how AI models are delivered. Accessing a model through a hosted service based outside the UK, meaning the model stays on foreign servers and users interact remotely, does not count as “importation” or “possession” under the CDPA’s secondary infringement rules. Downloading a model into the UK could count as importation, but only if the downloaded model was an infringing copy.

In this case, it wasn’t. But the distinction is now firmly established. For AI companies, this is a very big deal: remote access carries far less copyright risk than distributing model weights.

The elephant the court didn’t touch

For all the noise around this case, here’s the uncomfortable truth: the judgment does not answer whether UK-based web scraping and model training infringe copyright. Getty dropped the claims that would have forced the court to confront that foundational issue. As a result, the real legal fight, the one that determines the future of AI development in the UK, is still waiting in the wings.

This also means the UK Government’s ongoing struggle with AI copyright policy remains unresolved. With more than 11,500 consultation responses and mounting pressure from creative industries, the calls for legislative reform will only grow louder.