AI is a hot topic (and will be for the foreseeable future no doubt). Whether it’s the benefits, risks, ethics or impact on certain jobs – there are myriad talking points.
Here, we are looking at the landscape following one of most significant legal assessments yet of how copyright law applies to generative AI (Getty Images v Stability AI [2025] EWHC 2863 (Ch)).
Getty Images v Stability AI was a vast case, and the trial alone was 12 days long. It morphed over the two and half years between the case being issued and the final trial, with numerous hearings along the way, and during the trial itself significant parts of the claim were dropped by Getty Images.
At its core, the case was about the use of images, said to have been ‘scraped’ from the internet, to train an AI image generator (called Stable Diffusion), and whether the resulting model itself was effectively a copy of its training data. Getty Images’ case included a range of claims: primary copyright infringement, secondary copyright infringement, database rights, trade mark infringement and passing off. Of those, only secondary copyright infringement and trade mark infringement ultimately needed to be decided by the judge.
Whilst the question of whether training an AI model on copyright works was an infringing act has not been decided under UK law, the case has made clear the importance of establishing where training takes place to determine the proper forum for a claim and what law applies. This gives rise to the possibility of AI developers shopping around for jurisdictions with the most favourable laws on the issue.
The upshot was the primary copyright infringement fell away at trial, having previously survived an attempt by Stability AI to have it thrown out back in 2023.
Secondary infringement – a weighty issue
With primary infringement not needing to be decided, the most significant part of the judgment concerned secondary infringement. The essence of the claim was whether Stable Diffusion was an ‘article’ within the meaning of the Copyright, Designs and Patents Act 1988 and whether it was itself an infringing copy of Getty Images’ works. Then it would have to be decided whether Stable Diffusion had been imported into the UK.
Ultimately the secondary copyright claim was dismissed. It appears there was a significant amount of time spent dealing with the detailed technicalities of how Stable Diffusion was trained and functions, and to what extent the training dataset forms part of the model in operation. That makes for a (potentially) interesting read if you want some background on how AI is developed and functions.
The short version of the decision is that although an AI model could be an ‘article’ and therefore potentially within the scope of secondary infringement, Stable Diffusion did not actually store or reproduce copyright protected works and so could not be an infringing copy.
What Stable Diffusion does is use ‘model weights’ (essentially parameters of the model) which are abstract numerical patterns learned during training – they start off randomised and are then optimised based on the training data. These model weights do not correspond to discrete image files or stored copies of Getty Images’ works. The model does not get asked a question and then go and pull bits from various images to create a composite image of some description (it can’t, as the model as deployed does not store or access the training dataset).
As the judge put it, in simple terms you end up with ‘an input-output system in which the user specifies inputs, the trained network performs computations on those inputs and then generates the desired output’.
The issue became somewhat moot as Stable Diffusion was found not be an infringing copy, but it was said that the downloading of the model into the UK can be importation.
Getty Images also pursued trade mark infringement claims, alleging that Stable Diffusion generated images that carried its trade marks (as watermarks). Strictly speaking, Getty Images partially won this point, but it was a very narrow finding. It doesn’t really have wider application because it was based so heavily on its facts, but it does at least give some authority to the point that, in some circumstances at least, an AI model generating an existing trade mark could be an infringement of that trade mark.
So where does the law stand now?
With such novel and important issues at stake, there will almost certainly be attempts to appeal some of the findings – so the position could yet change.
As it stands however, AI developers and providers will be buoyed by this decision as it means that the provision of models such as Stable Diffusion into the UK are very unlikely to result in copyright claims.
The question of whether training an AI model in the UK is an infringement remains very much open. Battle lines are being drawn globally on all these issues, with some decisions coming out in the US, and the ongoing case of The New Your Times v OpenAI in US likely to have huge impacts given not only its scope but that it will cover one of the major jurisdictions for AI development and training. There is also an EU decision pending in Like Company v Google Ireland that covers copyright issues and AI.
All of that leaves aside the possibility of legislative or regulatory changes that might again shift the legal landscape. So, whilst we can say that answers are slowly emerging to the big legal questions around AI, it is far from a settled or certain position.
By Stephen Ruse, Director of Intellectual Property & IT Law at Howes Percival.
For more information, contact Stephen Ruse on 01604 230400 or 01908 672682 or
email stephen.ruse@howespercival.com.
For more about Howes Percival, visit their website.
