Getty Images v Stability AI: technical knockout on copyright

The High Court has handed down its much-anticipated decision in Getty Images (US), Inc & others v Stability AI Ltd, marking the first substantive UK ruling on whether the training of a generative-AI model infringes copyright.

The dispute

Getty (a global image licensing company) brought proceedings against Stability AI (the developer of the image-generation AI model Stable Diffusion), alleging that the company had copied millions of Getty’s photographs without permission when training the model.

Its claims spanned copyright infringement, database-right infringement, trade mark infringement (arising from the appearance of Getty and iStock watermarks in generated images), and passing off.

Stability AI denied liability, arguing that training had taken place outside the UK, that the model itself was not a copy of any protected work, and that the outputs were generated independently in response to text prompts.

What the court decided

In a judgment handed down on 4 November 2025, Mrs Justice Joanna Smith dismissed the majority of Getty’s claims.

Getty had originally advanced a ‘Training and Development Claim’ alleging primary copyright infringement during the model’s training. The Court accepted that Stability AI had used large datasets that included Getty material, but that claim was abandoned after Getty had failed to establish that any act of infringement occurring within the UK. Training and model development took place overseas, meaning the alleged copying fell outside the territorial scope of UK copyright law. Getty’s database-right claim also fell away.

The remaining secondary copyright claim related to the model itself.

Mrs Justice Smith accepted that an ‘article’ for the purposes of the CDPA can, in principle, be intangible.

However, the Court rejected Getty’s argument that the Stable Diffusion model itself constituted an ‘infringing copy’.  While the model’s parameters were shaped by exposure to copyrighted works, the judge found that “the Model itself does not store any of those Copyright Works; the model weights are not themselves an infringing copy and they do not store an infringing copy”.

The Court’s view was that as the model functions as a statistical system of weights and numbers (“model weights” being the numerical values learned by the system during training, which define how it generates images but do not store or contain the original photographs), the model itself was not an infringing copy of any Getty work. As a result, the question of importation or dealing under sections 22-23 did not arise.

The Court found limited success for Getty under the Trade Marks Act 1994.

Under section 10(1) (double identity) and section 10(2) (likelihood of confusion), Getty succeeded in respect of iStock marks appearing in images generated by early versions of Stable Diffusion.

All other trade mark claims (including those under section 10(3) (reputation/unfair advantage), and those concerning later versions of the model), were dismissed.

The judge declined to address passing off separately, given the partial success under trade mark law.

The overall result was therefore a narrow success for Getty, confined to specific marks and early versions of the model, with the vast majority of claims dismissed.

What if the training had occurred in the UK?

The judgment stops short of deciding whether the training of a generative-AI system amounts to “copying” under the Copyright, Designs and Patents Act 1988, because that claim could not be pursued.

But Mrs Justice Smith’s reasoning gives some indication of how such a claim might have been treated if the evidence had shown UK-based training.

Her discussion of the ’infringing copy’ claim makes clear that she regarded the training process and the resulting model as conceptually distinct. She accepted that the process of training involves the reproduction and storage of the copyright works in local and cloud environments, but emphasised that those acts were separate from the later existence of the trained model.

Had Getty been able to prove that those reproductions took place in the UK, the Court might therefore have treated them as acts of primary infringement under section 17; much as copying into RAM or temporary storage has been recognised as copying in earlier UK cases.

However, the judge also signalled caution about equating machine-learning processes with conventional copying. She described Getty’s suggestion that the model became an infringing copy as soon as it is made as “entirely misconceived”, and noted that the model “does not store any of those Copyright Works”. That language suggests a narrow approach: training might involve copying only where identifiable reproductions of protected works are stored or retained, not merely because the system’s parameters are influenced by them.

If the training had occurred in the UK, therefore, the outcome might have turned on how the data had been handled. Systematic downloading or storage of Getty’s images during training could plausibly amount to infringement; ephemeral processing that left no recoverable copy might not. The Court’s focus on storage and reproducibility implies that the line would be drawn at the point where a work (or a substantial part of it) is actually fixed in a medium from which it can be retrieved.

The judgment therefore leaves open the boundary between lawful learning and infringing reproduction.

What happens next

Getty has said it is considering its options, including appeal.

In parallel, related claims continue in the United States, where discovery rules and fair-use principles differ.

For now, UK developers gain some reassurance that, on the present law, training conducted abroad, and the resulting models, are unlikely to infringe UK copyright.

For right-holders, the decision signals that litigation strategy will need to focus on where copying occurs and what the AI system actually reproduces, rather than on the mere fact of training itself.

Key takeaways 

1. Territorial limits remain decisive
   The decision confirms that the UK’s copyright regime is still grounded in acts occurring within its borders. Right-holders will struggle to pursue claims unless they can demonstrate copying or distribution in the UK itself. 

2. The model is not a “copy”
   The judgment’s finding that Stable Diffusion’s model weights do not reproduce any work is fundamental. Training may involve exposure to protected material, but the resulting model is a set of mathematical parameters, not a library of images. 

3. Limited trade mark success for right-holders
   Getty succeeded only in narrow circumstances under sections 10(1) and 10(2). The findings turn on early versions of the model and specific examples of watermarks, offering little precedent for broader infringement claims. 

4. Policy questions remain open
   The Court left unanswered whether the training of AI models on copyrighted works should itself be regulated or licensed. That issue now lies squarely with policymakers.

Humphreys Law are specialists in copyright and trade mark licensing and enforcement.  If you have any questions or issues concerning the above, contact a member of our team.  This piece was written by Tristan Morse.

All the thoughts and commentary that HLaw publishes on this website, including those set out above, are subject to the terms and conditions of use of this website.  None of the above constitutes legal advice and is not to be relied upon.  Much of the above will no doubt fall out of date and conflict with future law and practice one day.  None of the above should be relied upon.  Always seek your own independent professional advice.

Humphreys Law