On 4 November 2025, the UK High Court handed down its highly anticipated judgment in Getty Images (US) Inc. v Stability AI Limited[1], the first UK judgment considering substantive issues of copyright and trade mark infringement in relation to AI-development and use. This alert summarises the key points of the judgment and identifies practical takeaways for rightsholders and other stakeholders in the AI value chain.
Key Takeaways
- Watch this space. The judgment ultimately gave very little clarity on the legality of AI models being trained on copyright works. Jurisdictional and evidentiary issues resulted in a very narrow role for the judge, and the decisions on secondary copyright infringement and trade mark infringement were highly fact-specific. As such, pending appeal, eyes will turn to UK Government consultations in this space and concurrent litigation taking place in the US and EU.
- Models are “articles”, but the weights may not be an “infringing copy”. The judgment is definitive that an AI model is an “article” for the purpose of UK copyright law. However, whether the model as an article, and more importantly the model weights, can amount to an infringing copy of copyright works appears to be fact-specific and with room for future challenges. In this case the Stable Diffusion model was held not to store its training data when made available in the UK, but that was specific to this model, and there were underdeveloped arguments regarding the ability of models to memorise by overfitting, which may have made the judge’s decision more complex. More clarity will be required on this point in future cases.
- Importance of safeguards. By taking proactive steps to filter certain prompts, training data and potentially infringing outputs, Stability was able to reduce the allegedly infringing acts and mitigate potential liability. Had it improved its filters for the Getty watermark, then the liability at first instance may have been nil. This highlights the importance of taking proactive steps to implement technological filters and other processes to address instances of actual or potential infringement, something many general purpose AI developers with models in the EU market will already be doing as part of their EU AI Act obligations.
Summary
Getty Images (US) Inc. (“Getty”) licenses millions of images curated over decades to users globally from a database which is publicly displayed on its website bearing watermarks containing “GETTY IMAGES” or “ISTOCK” trade marks. Stability AI Limited (“Stability”) is an AI developer, mainly developing large language diffusion models that generate images from text or image prompts, with its primary text-to-image model, “Stable Diffusion”, made public in the UK in 2022 (and updated thereafter).
In January 2023, Getty issued proceedings against Stability in the High Court, alleging primary and secondary copyright infringement, database right infringement, trade mark infringement and passing-off, based on Stability’s alleged use of Getty images to train the various Stable Diffusion models and its use of the Getty trade marks in outputs provided to users in the UK. By the end of the trial, Getty ultimately abandoned its claims for primary copyright infringement and database rights infringement on the basis of a lack of evidence that training and development of Stable Diffusion took place in the UK and given that Stable Diffusion was no longer capable of the infringing acts forming part of the claim. As such, the High Court ultimately answered the following two substantive questions:
- Does Stability’s release of its Stable Diffusion model in the UK amount to secondary copyright infringement of Getty’s images by importation of an article representing an infringing copy of Getty’s works with requisite knowledge by Stability?
- Does Stable Diffusion’s output of watermarked images, including the Getty trade marks, constitute trade mark infringement?
Secondary copyright infringement
The following two key aspects of the secondary copyright infringement test were considered:
- Meaning of “article”. The court found that the word “article” for the purpose of UK copyright law includes intangible property such as an AI model on the basis of: (i) related statutory terms expressly referring to intangible and electronic mediums and concepts such as “copying” referring to storing a work “in any medium by electronic means”; and (ii) the previous UK and EU case law that supported that conclusion.[2]
- Model weights as an infringing copy. On the basis that the parties agreed that Stable Diffusion’s model weights do not store visual information of any training data (including Getty’s images) and that reproduction of a work is critical for infringement, the Court found that the Stable Diffusion model is not an infringing copy of Getty’s images in the UK (given that the images are not within the model weights upon import into the UK).
Therefore, Getty’s secondary copyright infringement claim was dismissed, although before doing so the judge noted that: (i) there was evidence that Getty’s works were used to train Stable Diffusion (likely in the US); (ii) Stability’s making available of the model weights for download, and subsequent actual downloading, in the UK amounted to “importation”; and (iii) if Stable Diffusion had amounted to an infringing copy, Stability knew or had reason to know this fact (including based on several internal exchanges discussing the prevalence of the Getty watermarks).
Trade mark infringement
Despite the lengthy judgment on trade mark infringement, the judge’s analysis was fairly straightforward as a matter of law in holding that Stability had committed trade mark infringement (as a result of both identicality and likelihood of confusion). The key points of interest were that: (i) the average consumer was limited to users of Stable Diffusion (and not recipients of outputs shared); (ii) the output by Stable Diffusion of images bearing the Getty trade marks constituted a commercial communication by Stability of the Getty trade marks, which created the impression of a material link between such generated images and Getty; (iii) the user of Stable Diffusion is not in control of the model’s outputs by virtue of the prompt that they input in compliance with Stable Diffusion’s terms of use; and (iv) Getty’s claims in relation to dilution, tarnishment and unfair advantage of reputation of the Getty trade marks could not succeed because there was “not a scrap of evidence” of any change in economic behaviour by users as a result of Stable Diffusion generating watermarked images. Ultimately, Getty was able to prove that at least a limited number of watermarked images had been generated by at least some versions of its model, although the Court noted that given the difficulty in adducing evidence, the volume of such infringing watermarks in the UK is “unknowable”.
Comment
After much anticipation, the Getty case has failed to deliver any real substance or steer to assist rightsholders or AI developers. Without a decision on how training of AI models will be viewed in an analysis of primary copyright infringement, and with fact‑specific findings on secondary copyright infringement, there is limited indication of whether and how AI models can train using third-party content. Similarly, it remains to be seen, for example, whether AI models could be considered to “memorise” an exact image in its weights through overfitting (discussed as part of the trade mark infringement claim), which may have been helpful in determining that the model was indeed an infringing copy.
Whilst this decision is likely to be appealed, there may be a limit to how far the “always speaking” principle (i.e., that a statute should be interpreted taking into account changes since its enactment) can be applied considering that most UK IP laws were not drafted with the GenAI world in mind. Eyes will turn to the UK government to see how, following its consultation on copyright and AI[3], it will balance the interests of big tech on the one hand and the UK’s highly prized creative industries on the other.
Critically, the case implies that primary infringement of UK copyright works could be evaded by training models in other jurisdictions. In that context, the outcome of the ongoing parallel dispute between the parties in the US[4] (which is where the evidence implies that Stable Diffusion models are primarily trained in the US) may be more determinative. If the US courts ultimately find that training of a model via third-party content amounts to infringement, developers may be left scrambling to find a favourable jurisdiction to undertake those activities. Similarly, the ongoing referral from the Hungarian courts in the EU[5] regarding the legality of training AI models on copyright content in light of the EU text and data mining exemption adds more complexity to the global picture (though judgment is not expected until 2027).
* * *
[1] Getty Images (US) Inc. and Ors v Stability AI Limited [2025] EWHC 2863 (Ch), available here.
[2] Sony Computer Entertainment Inc v Ball [2004] EWHC 1738 (Ch), available here (in which a RAM chip was an article for the fraction of a second in which it stored an infringing copy of digital data); and (ii) Austro-Mechana Gesellschaft Zur Wahrnehmung mechanisch-musikalischer Urheberrechte GmbH v Strato AG (C-433/20) EU:C:022:217, available here (in which a copy of a work saved in the cloud was considered a reproduction of that work).
[3] Consultation on Copyright and Artificial Intelligence (17 December 2024 to 25 February 2025).
[4] Getty Images (US), Inc. v Stability AI, Inc. 3:2025cv0689, (D.Cal), docket available here.
[5] Like Company v Google Ireland Limited (Case C-250/25), case information available here.
Related Insights
