The much anticipated High Court judgment in Getty Images v Stability AI has been handed down. As we knew would be the case, with Getty having dropped some of its more crucial claims prior to closing arguments, it is not quite the trail-blazing precedent setter we initially hoped for, but there is still plenty to pick over.
At the outset we note the most relevant long-term point of applicability arising is the judge’s finding that the model was not an infringing copy and in particular the model itself did not store any of the copyright works. The exposure of model weights to infringing copies does not mean that the model itself is an “infringing copy.” The fact that its development involved the reproduction of copyright works (through storing the images locally and in cloud computing resources and then exposing the model weights to those images) is of no relevance. This a strong finding reducing the liability of AI models for copyright infringement.
Intriguingly, in a case that demonstrates the spectrum of judicial approach, a week later on 11 November 2025, in a very different case before the Munich Regional Court in Germany, the court found that ‘memorisation’ within a model was itself an act of reproduction and that where the ‘original’ elements of a copyright work (i.e. song lyrics) were rendered in a synthetic output that was a further reproduction and a communication to the public for which the AI model developer was responsible, rather than the AI model user.
What follows is our (very) high-level summary of what we believe are the three most important aspects, including the dropped claims, as what the court did not have to decide is equally interesting as means this is not the end of the issue.
1. The dropped claims
- Following the evidence and prior to submission of closing arguments, Getty dropped what we consider to be its most important claims. These included the following:
Claim | Comment |
Getty’s copyright claims regarding the use of Getty’s images in the training and development phase for Stable Diffusion | Getty had been unable to demonstrate that any relevant infringing activity had occurred within the UK. The issue of whether the use of copyright works to train and develop and AI model therefore remains unresolved before the UK courts. In its ongoing consultation on AI & Copyright the government has proposed a text and data mining copyright exception which is intended to permit training and development of AI models, subject to copyright holders being entitled to reserve their rights in this regard. |
Getty’s copyright claims that the generation of synthetic outputs by Stable Diffusion were infringing copies of Getty’s works | Getty suggested that its concerns had been largely met because Stability AI had applied filters to block the specific prompts which were alleged to have rendered the infringing outputs. A number of questions remain. For example, what level and degree of human input is required for a synthetic output to be considered a copyright work? In the AI & Copyright consultation referenced above, the government proposed to abolish section 9(3) of the CDPA which deems who the author is for computer-generated works. This is because the provision does not really work – a computer-generated work is one where there is no human author, but if there is no human author then it is difficult to see where the “intellectual creation” required for copyright subsistence comes from. |
Getty’s database right claims | These essentially mirrored the two copyright claims above. |
- This left Getty with, essentially, two main claims – namely, secondary copyright infringement and trade mark infringement.
2. The secondary copyright infringement claims
- For its claims of secondary copyright infringement, Getty was required to demonstrate that Stable Diffusion was an “article” that was an “infringing copy” and had been imported into the UK and/or possessed in the course of business and/or offered and/or distributed in the course of business, which Stability knew or had reason to believe was an infringing copy of Getty’s copyright works. Getty’s claim depended upon two key points:
Issue | Decision |
Was Stable Diffusion an article where an intangible copy of it had been downloaded in the UK? | The court found that an intangible copy of Stable Diffusion was an “article” and therefore would potentially fall within the provisions on secondary infringement under the Copyright Designs and Patents Act 1988. The court’s view was that the concept of an article could not be divorced from the concept of an infringing copy or indeed what amounted to a copy or copying under the CDPA 1988. Copying includes storing the work in any medium by electronic means. The court’s view was that parliament must have intended this to include storage by modern (intangible) means such as cloud storage. The court also relied upon the “always speaking” principle (that legislation should be interpreted taking into account changes that have occurred since it was enacted). This conclusion is perhaps unsurprising given the ubiquity of copyright works which are stored in electronic and intangible form due to advancements in technology. The finding is also likely to have wider relevance (e.g. in IT disputes). |
Was Stable Diffusion an infringing copy? | The court decided that Stable Diffusion was not an infringing copy, meaning that the claims failed. The judge noted that the model weights of Stable Diffusion were altered during training by exposure to copyright works but “by the end of that process the Model itself does not store any of those Copyright Works; the models weights themselves are not an infringing copy and they do not store an infringing copy”. The models are capable of reconstituting inputs (or substantial parts of inputs) within synthetic outputs. However, this is a result of “memorisation” rather than the model containing copies of those works on which they have been trained. |
3. The trade mark claims
- The court concluded that the question of trade mark infringement was limited to certain versions of Stable Diffusion which generated synthetic outputs featuring the relevant watermarks which were alleged to be infringing.
- The court drew a distinction between two different types of average consumer: (1) those that accessed Stable Diffusion from Stability’s developer platform API, who were technically savvy but who might assume there is a connection between Getty and Stability as a result of the appearance of the watermarks; (2) those who access Stable Diffusion through DreamStudio who were less technically knowledgeable and who may believe that the synthetic output featuring the watermarks were provided by Getty itself.
- Perhaps the most interesting aspect of the court’s decision on the trade mark infringement claims was the judge’s finding that Stability was responsible for use of the trade mark rather than the user of the model. The judge found that the watermarks rendered in the synthetic outputs would be perceived by the average consumer as a commercial communication by Stability. This was because the watermarks were affixed to synthetic outputs as a result of the functionality of Stable Diffusion and that was dependent upon its training data over which Stability had absolute control and responsibility. While this aspect of the decision was heavily fact-dependent, it suggests that AI model developers may be liable for trade mark infringement where infringing signs are generated in synthetic outputs.
- The trade mark infringement claims were decided as follows (the court declined to address Getty’s passing off claims, which stood or fell with the trade mark infringement claims):
Trade mark in question | Infringement claim | Outcome |
iStock marks | S10(1) TMA 1994 Use of an identical sign in relation to identical goods/services
| Getty succeeded on a limited and historic basis for iStock watermarks generated by users of v1.x of Stable Diffusion when accessed via Dream Studio and/or the Stability’s developer platform API. |
Getty Images marks | These claims were dismissed due to lack of evidence of infringement in the real-life examples Getty had produced, which were not identical. | |
iStock marks | S10(2) TMA 1994 Use of an identical/similar sign in relation to identical/similar goods/services
| Getty succeeded on infringement for iStock watermarks generated by users of v1.x of Stable Diffusion. |
Getty Images marks | Getty succeeded on infringement for Getty Images watermarks generated by users of v2.x of Stable Diffusion. | |
iStock and Getty Images marks | S10(3) TMA 1994 Use of an identical/similar sign where the trade mark has a reputation which without due cause takes unfair advantage of or is detrimental to the distinctive character or repute of the mark | These claims were dismissed. There was no evidence that there would be continuing use of the watermarks that would cause detriment to the distinctive character of Getty’s marks. Neither was there evidence of images bearing the watermarks which would cause detriment to Getty’s reputation. There was no basis for inferring that Stability intended to take unfair advantage of Getty’s marks. |
- The findings of trade mark infringement were very limited and arguably no more than a Pyrrhic victory for Getty.
/Passle/5f3d6e345354880e28b1fb63/SearchServiceImages/2025-11-14-12-10-46-861-69171c468b283f3b3abda930.jpg)
/Passle/5f3d6e345354880e28b1fb63/SearchServiceImages/2025-11-14-11-58-28-980-691719648b283f3b3abda02e.jpg)
/Passle/5f3d6e345354880e28b1fb63/SearchServiceImages/2025-11-13-17-15-02-144-69161216e074526d10a5bc45.jpg)