7 minute read | November.18.2025
The UK High Court’s judgment in Getty Images v Stability AI marks a significant development in how English law approaches generative AI models, particularly latent diffusion systems like Stable Diffusion. As Mrs Justice Smith succinctly emphasised the importance of this judgment in her opening remarks “Both sides emphasise the significance of this case to the different industries they represent: the creative on the one side and the AI industry and innovators on the other. Where the balance should be struck between the interests of these opposing factions is of very real societal importance.”
The Court addressed whether AI models can be “articles” capable of being “infringing copies” under the Copyright, Designs and Patents Act 1988 (CDPA), and examined trade mark claims arising from AI-generated outputs.
While the Court acknowledged that intangible electronic items can be “articles,” it held that latent diffusion model weights are not “infringing copies” of training images. On trade marks, the Court found limited instances where watermarks in AI-generated outputs infringed trade marks, though largely such infringement was constrained to particular model versions, access pathways, and watermark types, stressing that infringement was a case-by-case analysis.
The decision provides guidance on the nuanced application of copyright law to training AI models and trade mark law to AI outputs, as well as territorial limits of UK copyright law and evidential burdens in secondary infringement cases.
Getty Images supplies stock photos, video, and other content for commercial use. Stability AI, a UK-based company, developed Stable Diffusion, an open-source image generative model. Getty brought proceedings in the UK against Stability AI alleging copyright infringement (primary and secondary), database right infringement, trade mark infringement, and passing off. Similar claims were made in parallel US litigation.
Late in the trial, Getty narrowed its case, abandoning primary copyright infringement and database right claims:
The case proceeded on secondary copyright infringement under CDPA ss. 22–23 and the definition of “infringing copy” in s. 27, and on trade mark infringement under the Trade Marks Act 1994 (TMA).
Stable Diffusion is a latent diffusion generative model. It transforms prompts into synthetic images by modelling probability distributions learned from training data and sampling from those distributions. Model parameters (weights and biases) define the network connections in the model and are learnable parameters controlling the functionality of the network. These parameters are initialised with random values and then optimised through exposure to the training data. Once trained, the network generates outputs without reusing the training data known as “inference”. The Court opined that “inference does not require the use of any training data and the model itself does not store training data. However, a large part of its functionality is indirectly controlled via the training data”..
The Court accepted expert evidence that the model does not store pixel-level copies of training images; rather, it encodes patterns in a compressed latent space. The experts did agree, however, that the model can produce synthetic outputs that are distinct from the training examples, it can also produce images that are derived from a training image either in part or in whole potentially as a result of memorization and overfitting occurring in the model.,
Getty argued that the model weights are an intangible “article,” and that their making via repeated exposure to copyrighted works, would constitute infringement if trained in the UK. Stability countered that “article” refers to tangible items and, in any event, that the model does not contain copies of copyrighted works.
The Court held that “article” under ss. 22–23 can include intangible electronic items; limiting protection to tangible objects would undermine enforcement where copying is electronic. However, the Court rejected the idea that Stable Diffusion’s weights are an “infringing copy” under s. 27. Even if training involves exposure to infringing copies, the resulting model weights do not store those works; they embody learned patterns. Therefore, the model is not itself a reproduction of the copyrighted works, and the secondary infringement claim failed.
Getty presented engineered “Getty Watermark Experiments,” designed by Getty’s lawyers and their expert witness, showing that particular prompts can elicit outputs featuring watermark-like artefacts. While Stability accepted that certain models could be “pushed” to generate such artefacts, it questioned the probability or frequency of such outputs in real-world use. Both parties’ experts ultimately agreed the experiments likely overestimated the prevalence of these outputs due to prompt bias from using Getty captions as input, and that later models employed filters that reduced (though not eliminated) the likelihood of watermark reproduction..
The Court considered not only the differences between model versions but also varying access pathways and the resulting impact on the extent to which the experiment prompts could have been used by real world users to obtain the same output. Not all versions were trained on identical datasets, and filters varied. As a result, capability to reproduce watermarks differed across versions.
The Court identified three relevant categories for the average consumer: local download users (technically skilled), API/developer platform users (technically capable), and DreamStudio users (less technical, using a web interface). A proposed fourth “bystander” class was rejected as speculative. The assessment of trade mark confusion and use in the course of trade for the purposes of determining trade mark infringement was tailored to these realistic pathways.
This judgement clarifies that while intangible electronic items can be “articles” under UK copyright law, the trained weights of a latent diffusion model are not “infringing copies” merely by virtue of being trained on data protected by copyright. It underscores UK copyright’s territorial limits and the evidential demands for secondary infringement claims, particularly regarding proof of UK-based acts and the nature of copying in AI models.
On trade marks, the decision finds that identifiable watermarks in AI-generated outputs can constitute use in the course of trade and, in certain circumstances, give rise to identical or confusingly similar use. However, infringement is case-specific, being highly dependent on model version, dataset filtering, access pathway, and the specific output. Courts will consider technical realities, such as memorisation risks and prompt engineering, when evaluating AI-related trade mark claims.
Overall, the judgment is a cautious, technically informed application of established legal principles to generative AI. It avoids broad pronouncements, instead offering targeted guidance that developers, platforms, and rights holders can use to assess risk and implement mitigations, including dataset curation, filters, and output monitoring.