Getty Images v Stability AI - Getty's narrow win in trade mark infringement claim
Judgment on the closely watched Getty Images v Stability AI case has been handed down today. Our previous article (see here) outlined the background to the case and the key issues to be considered by the Court.
We have summarised the Court’s finding as regards the various copyright claims in a separate article here, focusing in this article on Mrs Justice Joanna Smith’s findings relating to Getty’s claims of trade mark infringement and passing off.
Mrs Justice Joanna Smith found that limited infringements occurred under ss.10(1) and 10(2) of the Trade Marks Act 1994 in relation to specific and quite limited instances of watermarks generated by early versions of Stable Diffusion. Further, Getty’s s.10(3) reputation claim was dismissed and Mrs Justice Joanna Smith chose not to rule on the passing-off claim.
As a brief re-cap, you may recall that Getty brought UK proceedings asserting infringement of its registered trade marks, including the well-known GETTY IMAGES and ISTOCK marks, arising from Stability’s text‑to‑image model, Stable Diffusion. Getty alleged that Stability trained multiple versions of the model on large web‑scraped datasets that included Getty’s watermarked content without consent, and that, in normal use, users could task Stable Diffusion with generating synthetic images which reproduced the GETTY IMAGES and/or ISTOCK signs as watermarks, thereby using identical or similar signs to Getty’s registered trade marks. On that basis, Getty advanced claims under sections 10(1), 10(2) and 10(3) of the Trade Marks Act 1994, as well as passing off, relying in part on experiments demonstrating outputs bearing watermarks (for example, an AI‑generated scene featuring a visible “Getty Images” banner) to show both the occurrence of such signs and the likelihood of consumer confusion.
Section 10(1) – Double identity
The court considered legal tests for s.10(1) of the Trade Marks Act 1994 (double identity) in the following order:
- Was there “use” of a sign in the course of trade and “in relation to” goods/services;
- Was the sign and goods/services identical to Getty’s marks; and
- Whether trade mark functions were affected.
The Court ultimately held that s.10(1) (double identity) infringement was only found in the context of v1.x version of the Stable Diffusion model, and only in respect of the ISTOCK marks.
“Use” and “in relation to” goods/services
Getty had to prove Stability used the watermarked sign “in the course of trade” and “in relation to” the goods/services covered by Getty’s marks. “In the course of trade” means commercial use aimed at economic gain, not private use. Getty argued that, because Stability operates Stable Diffusion commercially in the UK and trained its model using watermarked images, it is Stability, not individual users, that procures the complained‑of watermarks in outputs.
The court held that on the facts, Stability’s role in the “use” of a sign went beyond purely passive enablement and that affixing a watermark sign to output images can constitute “in relation to”, for the purposes of the s.10(1) assessment.
Sign identity
In relation to identity of the marks, the Court concluded that the ISTOCK marks were identical to the alleged infringing marks in question, but found no real-world examples of the GETTY IMAGES marks being reproduced in identical form by Stable Diffusion.
Identity of goods and services
In this respect, the Court held that Stable Diffusion outputs constituted “synthetic image outputs” which may be identical to Class 9 goods (e.g., digital media/images), and that the provision of those outputs may be identical to Class 41 services (e.g., digital imaging services).
Court’s findings
For the ISTOCK marks, infringement was only found for v1.x of the model, whilst for the GETTY IMAGES marks, the claim was dismissed for v1.x and v2.x. The finding of s.10(1) infringement for the former appeared to turn on there being a specific real-world example of the watermark appearing in what was referred to during trial as the “Spaceships Image”.
Section 10(2) – Likelihood of confusion
Getty was partly successful in its claim for trade mark infringement under s.10(2), for specific models only. The court found s.10(2) trade mark infringement in the v1.x outputs (for the ISTOCK mark) and the v2.1 outputs (for the GETTY IMAGES marks). No infringement was found under s.10(2) for SD XL or v1.6, because there was no evidence of real-world UK generation of watermarked images. The judgment was careful not to generalise beyond the instances of proven infringement in this particular case.
Similarity of marks and goods/services
There was no evidence of Stability using an identical sign to the “Getty Images” sign. The watermark on the “First Japanese Temple Garden” images contained an extra “i” in such that it read “imaiges”, which defeated an identity finding for the purposes of s.10(1), but it was found highly similar for the purposes of s.10(2) (see para 415). The case indicates that where AI tools may blur marks in generated outputs, this may prevent a finding of identity.
The s.10(2) proceedings proceeded on a high degree of similarity only for specific specifications. The judge found that synthetic image outputs of “vector art” fall within the words “downloadable digital illustrations and graphics”. Stability also acknowledged that there was some similarity here.
Likelihood of confusion
The court conducted a global assessment of likelihood of confusion based on the average consumer, in particular taking into account the high degree of distinctiveness of the ISTOCK and GETTY IMAGES marks.
The judge recognised three type of average consumer (see para 265). These were (i) technically skilled users that downloaded and ran the model locally, (ii) technically skilled users accessing Stable Diffusion via the Developer Platform/API, and (iii) users of DreamStudio (who were less technologically sophisticated).
The judge found that where watermarks appeared clearly, a significant proportion of consumers would likely believe that there was a commercial connection between Getty and Stability (even for consumers that were technologically sophisticated, and understood why such watermarks arise). It was noted that platform branding and “garbled” watermarks (see image below) could reduce confusion, although it would not eliminate it.
The court also applied the Supreme Court’s guidance on post-sale confusion in the recent case of Iconix Luxembourg Holdings SARL v Dream Pairs Europe Inc [2025] UKSC 25 (UK Supreme Court), emphasising the role of context in assessing likelihood of confusion (see para 272), and considered the “realistic and representative way in which the average consumer will encounter the sign.” The Court held it was enough to see a watermark-like sign when the image was generated on a screen – it did not matter that the user had already chosen to use the AI tool.
Section 10(3)
By way of reminder, Getty had pleaded infringement under s.10(3) on the following bases:
- Unfair advantage (free-riding): Getty submitted that the benefit of the Getty trade marks, stemming from the power of attraction owing to the marks’ extensive reputation, will be transferred to Stability through their use of identical/similar signs.
- Detriment to distinctive character (dilution): While direct proof of a change in consumer economic behaviour is difficult, it can be inferred from the circumstances and the nature of the trade. Getty argued that users turned to Stable Diffusion to avoid licence fees, evidencing such a change.
- Detriment to reputation (tarnishing): Output including pornography, violent imagery and propaganda bearing Getty-style watermarks risked suggesting a connection with Stability and harming Getty’s reputation for high-quality content. Getty also sought to rely on alleged CSAM, but after expedition to the Court of Appeal, those allegations were held insufficiently pleaded and excluded from trial. Getty further argued that applying signs to synthetic, non-photographic output erodes the marks’ guarantee of authenticity.
Mrs Justice Joanna Smith rejected each limb of the above. On unfair advantage, the Court found no evidence of any intention by Stability to exploit Getty’s reputation, nor any indication that consumers sought images bearing watermarks. On dilution, the Court rejected the hypothesis that users turned to Stable Diffusion to avoid licence fees and held there was no proof of any change in the economic behaviour of the average consumer. On tarnishment, although Stable Diffusion can generate “not suitable for work” (NSFW) outputs, there was no real‑world UK evidence of Getty‑style watermarks appearing on pornographic, violent or deepfake content, and the alleged reputational harm from poor‑quality or distorted watermarked synthetic images was considered speculative absent evidence of actual market impact.
Passing off
In relation to passing off, Getty alleged the “classical trinity” of goodwill, misrepresentation (by the use of the Getty trade marks in output) and damage was present. Stability disputed the presence of misrepresentation, referring to multiple arguments including that:
- users would have read the T&Cs and will appreciate that the watermarks are a result of their own image synthesis;
- users will acknowledge the presence of watermarks is caused by Getty’s watermarks being used in training data; and
- the watermarks’ appearance is not consistent with the conventional branding.
Stability concluded that Getty had not shown evidence of damage.
The court noted that Getty’s case “stood or fell” with its trade mark claims and that any passing off claims added nothing to the narrow findings of trade mark infringement under ss.10(1) and 10(2) anyway. Accordingly Mrs Justice Joanna Smith chose not to determine the passing off claim.
Our thoughts
Consistent with the court’s own characterisation of its trade mark findings as “both historic and extremely limited in scope,” the judgment highlights the court adopting a cautious, fact‑specific approach. While the outcome will be encouraging for AI developers, it underscores the continuing need for rigorous output monitoring to ensure that generated content does not inadvertently reproduce identical or highly similar third-party marks, and for robust mitigation measures to reduce the risk of trade mark infringement or passing-off.
The authors would like to thank Natalia Martosinska, Trainee Solicitor at CMS, for her assistance in writing this article.
