Getty Images v Stability AI Judgment
The High Court handed down its judgment in Getty Images v Stability AI on 4 November 2025, a case widely regarded as the first and most important case in the UK which dealt with the intersection of generative AI and intellectual property law. While the scope of the dispute narrowed significantly towards the end of the trial because Getty abandoned several claims, the decision nonetheless addresses questions about the application of IP rights in the age of GenAI. This article focusses on the trade mark issues.
Secondary Copyright Infringement Rejected
The court dismissed Getty’s claim that Stability’s UK activities amounted to secondary copyright infringement. Training of the Stable Diffusion model occurred overseas, and its subsequent importation and distribution in the UK did not constitute dealing in an “infringing copy”. However, following the judgment, on 16 December 2025, Mrs Justice Joanna Smith granted Getty permission to appeal this to the Court of Appeal
Limited Trade Mark Infringement Found
Getty succeeded on a narrow point based on historic and extremely limited instances of trade mark infringement under sections 10(1) and 10(2) of the Trade Marks Act 1994 (‘TMA’), arising from watermarks appearing in certain outputs. The court rejected claims under section 10(3) and declined to rule on passing off. On 16 December 2025, Mrs Justice Joanna Smith refused to grant permission to Stability AI to appeal her findings on trade mark infringement to the Court of Appeal.
Trade Mark Issues: A Closer Look
Getty uses watermarks in its images which include the words ‘GETTY IMAGES’ or ‘iSTOCK’ which have been registered as trade marks for various goods and services.
The TMA sets out three circumstances under which trade mark infringement can occur and which is assessed by reference to the ‘average consumer’ or the ‘relevant public’.
Section 10(1) prohibits use of a sign identical to a registered trade mark in relation to identical goods or services.
Section 10(2) prohibits use of a sign which is identical or similar to a trade mark and the goods or services are identical or similar to those covered by the trade mark and where there is a likelihood of confusion.
Section 10(3) focusses on the protection of the trade mark owner’s reputation and applies in cases where the goods and services are not identical but use of the sign takes unfair advantage of the trade mark reputation (free-riding), causes detriment to the distinctive character (tarnishment) or repute (dilution).
Getty’s Trade Mark Claim
Getty Images claimed that Stability AI infringed its trade marks under sections 10(1), 10(2) and 10(3) of the TMA. The allegation centred on the appearance of watermarks, argued to be identical or similar to Getty’s registered marks which Getty claimed were present in the synthetic image outputs generated by the Stable Diffusion model.
Getty argued that these watermarks constituted use of a sign in relation to services, namely the provision of synthetic image outputs. While acknowledging that not every output contained such watermarks, Getty maintained that it had produced evidence demonstrating that at least, some outputs did.
Getty claimed that watermarks used were identical or similar to its registered marks: GETTY IMAGES and iSTOCK, which appeared in outputs generated by Stability’s Stable Diffusion model.
To succeed, Getty had to establish that at least one infringing output was generated in the UK. The court undertook a model-specific review and concluded that there were “historic” and “extremely limited” instances of infringement under sections 10(1) and 10(2). The claim under section 10(3) was rejected.
The Court’s Assessment
The Court held that there were at least three classes of average consumer. First, those who use the downloadable mode, second, those who use the developer model and third, those who use the web-based model depending on the technical skill.
The court assessed infringement on a version-by-version basis of Stable Diffusion. This granular approach reflected the technical reality: Stability’s models were trained on different datasets and applied varying filters over time. According to the evidence presented at trial, a filter was designed by Stability for its later models, to exclude the potential for the production of watermarks in the synthetic images, by removing watermarked images from the training dataset. The judge however, noted that not all models have been trained on the same dataset and that different filters have been applied.
Based on Getty’s evidence, the court held that that earlier versions of Stable Diffusion were more likely to produce outputs containing Getty watermarks and that later releases of Stable Diffusion were shown to be less likely to produce Getty watermarks as filtering became more efficient.
Getty relied on experimental prompts designed to elicit watermarked outputs. Stability argued these were contrived and unrepresentative of real-world usage, noting that most users would avoid generating images with watermarks. The court agreed in part, finding that prompts copied verbatim from Getty captions or loosely inspired by them were unrealistic. However, it accepted that certain generic prompts, such as “news photo” and “vector art”, were representative of actual user behaviour. Getty also presented examples generated “in the wild” using random prompts. While their UK origin could not be proven, the court treated them as probative of likely UK generation.
In relation to section 10(1), the Court considered that the provision of synthetic image outputs falls within class 41 under “digital imaging services” and “downloadable digital illustrations and graphics” with class 9 , and therefore, infringement. However, the Court rejected Getty’s claim that the provision of synthetic image outputs fell within “computerized on-line search and retrievable services for images…” in Class 42.
In considering section 10(2) the judgment explored whether post-sale confusion should be considered, following the Supreme Court’s guidance in Iconix v Dream Pairs earlier last year. In making the global assessment leading to a likelihood of confusion on the part of a significant proportion of the relevant public, court considered that it does not matter that the watermarks would only be encountered in a post-sale context. The court considered all the evidence including evidence that many consumers will be familiar with use of watermarks to mean that content must be paid for, and held infringement under section 10(2) for limited outputs .
The Court’s Decision
The court held that Stability’s conduct amounted to “use in the course of trade”. While users had some control over prompts, they could not control the watermark appearance. The court considered the inclusion of watermarks was the result of Stability’s design choices, not user intent.
Getty’s arguments based on section 10(3), on dilution, tarnishment and free-riding were dismissed. The court found no evidence of detriment to distinctive character or reputation, nor any unfair advantage. Crucially, Getty failed to demonstrate a change in economic behaviour of the average consumer or a serious likelihood of such change.
The Impact of the Judgment
The judgment clarifies that trade mark infringement in the GenAI context will turn on particular facts such as the realistic nature of user prompts and the nature of user control. Given that the trade mark aspects of this case concern a watermark artefact, unintended and unwanted by the user, it remains to be seen how courts will grapple with potential future cases where outputs contain trade marks in different contexts where brand owners may seek to challenge AI developers.
Where brand owners pursue trade mark infringement claims against an AI model developer, the authenticity of prompts and the consistency of allegedly infringing outputs may become relevant factors in assessing the claim. In cases where outputs vary in presentation or use of the mark, evidence may need to address each output individually.
