Author: Foreign Attorney Dmytro Izotov

The development of artificial intelligence (“AI”) technology continues at an unprecedented speed with new AI applications capturing the attention of millions of users worldwide. Although law is trying to keep pace with such rapid developments, currently there is an obvious lack of useful guidelines on the correlation between AI and intellectual property law. Against this background, recently the attention of the global legal community turned to the United Kingdom, where the High Court of England and Wales (“Court”) was tasked with deciding the first major AI-related dispute on British soil in case Getty Images (US) Inc and Others v Stability AI Ltd [2025] EWHC 2863 (Ch).

Below we present the details of the said case.

I. Outline of the Dispute

The Claimants, likely well-known to most our readers, are a prominent group of companies operating under the Getty Images and iStock brands and engaging in licensing of millions of images, videos, and other visual assets. Average internet user at least once stumbled upon photos with the Claimants’ distinctive marks when searching online. Over more than two decades, the Claimants built an immense library of content, earning the trust of creators and customers around the world.

In turn, the Defendant, Stability AI Ltd, was only founded in 2019. Despite being a relatively new company, it quickly became an influential player in the AI industry. Its flagship product, Stable Diffusion (“Model”), is a deep learning model capable of generating images from text prompts.

In early 2023, the Claimants issued proceedings in the Court alleging that the Defendant scraped and used millions of their copyrighted images without consent to train the Model. According to the Claimants, this not only infringed copyright but also caused the model to occasionally produce images displaying Getty Images or iStock-style watermarks. They further claimed that importing and distributing the Model in the UK constituted secondary copyright infringement.

II. Issues before the Court

The dispute initially raised several issues watched closely by businesses and IP practitioners worldwide:

(1) Whether training the Model on the Claimants’ content without consent constituted copyright infringement.

(2) Whether making the Model available in the UK constituted secondary copyright infringement.

(3) Whether AI-generated images bearing Getty Images/iStock watermarks amounted to trademark infringement.

AI Training

As you probably know, copyright law is territorial in nature. In the UK, under Section 57 of the Copyright, Designs and Patents Act 1988 (“CDPA”), infringement of copyright can only occur if the relevant acts take place within the UK. Therefore, to succeed on their training claim, the Claimants needed to prove that the Model was trained in the UK.

Understanding this, before the beginning of the trial, the Defendant applied for strike-out and/or summary judgment, arguing that all training activities occurred outside the UK with the use of cloud infrastructure located in the United States. While the Court declined to strike out the claim at that stage, the Claimants ultimately abandoned it before the trial. They accepted the lack of sufficient evidence to establish that any part of the training occurred within the UK.

Therefore, the proceedings focused on the remaining issues: secondary copyright and trademark infringements.

Secondary Copyright Infringement

Sections 22 and 23 CDPA impose liability where a person imports into the UK, possesses in the course of business, or distributes “an article which is, and which he knows or has reason to believe is, an infringing copy of the work.”

The parties disagreed on two questions: Was the Model an “article”? If so, was it an “infringing copy”?

The Claimants argued for a broad interpretation, submitting that “an article” may include intangible objects, like AI models, whose creation involves the use of copyrighted material. The Claimants insisted that the Model weights (the numerical parameters that a model learned to achieve its objective) exposed to the Claimants’ images during the training were such “infringing copies” in the present case.

The Defendant countered that only tangible objects can be “articles” within the meaning of the CDPA. Moreover, the Defendant stressed that the Model weights cannot be “infringing copies” because they do not store copyrighted works but only learn patterns and statistical correlations.

After analyzing the arguments from the parties, Mrs Justice Joanna Smith DBE held that intangible items can indeed fall within the statutory definition of an “article”. At the same time, she agreed with the Defendants on the key point that the Model was not an “infringing copy.” The judge elaborated that under the CDPA, a “copy” requires an act of reproduction of an article. In other words, such article must at some point contain or store the copyrighted work.

Although the Model weights were influenced by exposure to the Claimants’ images, they never stored those images or reproduced them in any form. Based on this, the judge concluded that the use of copyrighted images during training does not make the resulting Model an infringing copy, effectively dismissing the Claimants’ claim.

Trademark Infringement

The remaining issue concerned the occasional generation of images containing similar or distorted versions of the Getty Images or iStock watermarks.

Example of outputs generated by the Model (as reflected in the judgement):

The Claimants relied on Sections 10(1) (identical sign), 10(2) (likelihood of confusion) and 10(3) (protection of marks with a reputation) of the Trade Marks Act 1994.

With respect to Sections 10(1) and 10(2) of the Trade Marks Act, the Court found that several watermarks generated by the initial versions of the Model (including the image of the Japanese garden above) were similar to the Claimants’ marks and could give rise to a likelihood of confusion. By contrast, the Court determined that there was no evidence that the improved, subsequent versions of the Model produced any watermarks resembling the Claimants’ marks. Accordingly, the Court dismissed all allegations of trademark infringement as to the newer versions and concluded that any infringement was limited to just a few images generated by the early versions of the Model.

As for Section 10(3), the Claimants argued that the generated images bearing the watermarks harmed the reputation of their brand and allowed the Defendant to obtain an unfair advantage. The Court rejected this claim, holding that the Claimants had presented no evidence to support either reputational harm or the acquisition of an unfair advantage. In the Court’s view, identical or similar watermarks were observed only in a small portion of outputs produced by the initial versions of the Model, while most generated watermarks differed substantially from the Claimants’ mark and were unlikely to affect their reputation. There was also no evidence that consumer trust had been impacted in any way, nor any substantiation of the alleged unfair advantage. To the contrary, the Court noted that watermarked images were, in fact, undesirable for users, and that the Defendant had implemented a filtering mechanism in the next Model version to address the issue. This became another factor that supported dismissal of the Claimants’ Section 10(3) claim.

III. Conclusions

In today’s world, numerous issues arise concerning the relationship between AI and copyright. In our previous article, we explored the question of who can be protected by copyright in relation to AI-generated works from the U.S. perspective. This time, we examined another dimension of this relationship – the use of third-party copyrighted material during the process of generating AI content under the UK law.

Unfortunately, despite the judgment’s length and complexity, it ultimately provides only limited guidance. The Claimants’ decision to partially abandon their claim left unresolved the critical question of whether training AI models on copyrighted material within the UK constitutes infringement. Although the judgment had the potential to offer a framework that other countries could follow, this issue remains open for future litigation.

However, the Court did provide clarity on how AI models should be treated in the context of secondary copyright infringement. It adopted an approach likely to be welcomed by the AI sector, rejecting the idea that an AI model can constitute an infringing copy. This leaves copyright holders facing significant uncertainty about how to protect their works in the context of large-scale data scraping and model training.

With respect to trademark infringement, the judgment provided a road map for claimants wishing to succeed in similar claims. According to the Court, a claimant must present concrete examples of generated images that are identical to the trademark or sufficiently similar to give rise to a likelihood of confusion. Moreover, where a claimant alleges harm to reputation or the defendant’s unfair advantage, it must provide evidence that the generated images actually altered consumer perceptions or behavior toward the brand, and that the AI developer obtained a specific, demonstrable benefit as a result. Absent proof satisfying these requirements, trademark infringement cannot be established under the Trade Marks Act.

We will continue to keep you updated on new developments in the sphere of AI to see if courts of other countries may provide useful guidance on the related issues.