In the recent judgment in Getty Images v Stability AI [2025] EWHC 2863 (Ch), the High Court considered whether the generative AI model Stable Diffusion infringed copyright in works owned by/licensed to Getty Images, and further whether the model outputs infringed Getty Images’ trade marks. Getty argued that millions of its images had been used without permission to train the Stable Diffusion model, and that the model itself was therefore an infringing copy of the works.
Crucially, the court was not considering whether copyright was infringed during the training process of Stable Diffusion, as those claims were not pursued to trial by Getty due to a lack of evidence of training having been taken place in the UK. Instead, the High Court decided on the much narrower issue of whether the trained Stable Diffusion model is itself an “infringing copy” of the copyright works trained on. If the model was an infringing copy, under secondary copyright infringement law, its import into the UK would have infringed Getty’s copyright, even though the model had not been trained in the UK.
The High Court’s decision came down to the way in which Stable Diffusion was trained, and the relationship between the model and its training data. Stable Diffusion is a diffusion model, meaning its model weights are numerical parameters learned from training, not stored or compressed copies of its training data. The model does not contain any of Getty’s copyright images in any form whatsoever – and never has done – even though it may have been exposed to them during training. Getty’s secondary copyright claim failed as a result.
Although Getty lost its secondary copyright infringement claim, this was a highly fact-specific decision which related to this model of Stable Diffusion only. The High Court stressed this in its decision. Although it may be true that an “AI model which does not store or reproduce any copyright works (and has never done so) is not an “infringing copy””, this leaves the door open for an AI model that does store or reproduce copyright works (or has done so at some point) being found to be an infringing copy of its training data. Other model architectures that retain or reproduce their training data verbatim – which is more common for text models than image models like Stable Diffusion – may still be deemed infringing copies. In addition, there is scope for argument on whether a more liberal interpretation of what is an infringing copy should be adopted: in circumstances where the model has extracted the value and intellectual creation of copyright works, and in a manner that was not envisaged when the legislation was passed, why is this not reproduction of the underlying intellectual creation?
Further, as Getty dropped its training claims at trial, the UK courts are yet to decide on whether the training of AI models using copyright works in the UK infringes copyright. That question will need to be decided in a future claim involving an AI model that was trained (or at least partially trained) in the UK.
On the trade mark infringement claim, the court made a limited finding of trade mark infringement where early model versions of Stable Diffusion produced outputs with Getty-style watermarks.
If you’d like to speak to a member of the team about any of the issues raised by the judgment, please reach out to one of our AI experts.