UK court weighs in on AI copyright debate

 
Copyright is currently the hottest topic in AI litigation. Courts around the world are grappling with the task of mapping copyright frameworks onto the training and use of generative AI models. The England and Wales High Court of Justice weighed in on this discussion in Getty Images et al v. Stability AI Limited, 2025 EWHC 2863, an intellectual property dispute involving a stock photo company and an image-generating AI tool.

What you need to know

• The Court dismissed a claim for secondary copyright infringement, concluding that an AI model trained on copyrighted images is not itself a “copy” of those images for the purposes of the UK Copyright, Designs and Patents Act 1988.
• The decision did not address whether the training of the model constituted copyright infringement, as Getty Images dropped that issue after hearing evidence that no model training had taken place within the United Kingdom.
• The Court found limited instances of trademark infringement where the model had affixed watermarks to AI-generated images that could confuse users into associating Getty with the generated image.

The decision

Background

In January 2023, Getty Images brought a claim against Stability AI. The claim was initially broad, alleging that Stability’s AI model (Stable Diffusion) infringed copyright by (among other things) training on Getty’s photographs and producing infringing outputs. However, the claim narrowed considerably during trial. Getty dropped its primary copyright infringement claims because (a) there was no evidence that any training had occurred within the United Kingdom, and (b) Stability agreed to block prompts that had been used to generate potentially infringing output. At trial, Getty continued to advance secondary copyright infringement and trademark infringement claims.

Facts

Stable Diffusion was trained on images downloaded from an open-access dataset known as LAION-5B, which includes URLs, text captions and other information about publicly accessible online images. Stability led evidence that all downloading, storage and training associated with the dataset took place on an Amazon Web Services cluster located outside of the UK.

The LAION-5B dataset contained images from Getty’s online gallery. Many of these images were watermarked with a Getty or iStock logo.

No secondary copyright infringement

Getty claimed Stability infringed copyright by importing Stable Diffusion into the UK, arguing that Stable Diffusion was an “infringing copy” because its model weights were developed in a manner that would have constituted copyright infringement had the training occurred in the UK. The Court disagreed, finding that an “infringing copy” must be an actual copy of the works in question. Stable Diffusion was not a “copy” because it did not store training data, and there was no evidence that it had memorized copyrighted works.

Limited instances of trademark infringement

Getty alleged that Stability infringed its trademarks because Stable Diffusion generated watermarks resembling Getty and iStock watermarks to AI-generated images. The Court found that watermarks generated by certain versions of Stable Diffusion violated Getty’s rights under sections 10(1) and 10(2) of the UK Trade Marks Act, 1994. However, there was insufficient evidence that the distinctive character of the trademarks had been diminished under section 10(3). Further, there was no evidence that Getty’s reputation would be damaged by watermarks appearing on inappropriate content, nor that Stability had intended to take advantage of Getty’s reputation.

Impact

This decision has received significant attention as one of the early pronouncements on the intersection between AI and intellectual property law. But while “historic”, as the Court observed, it is also “extremely limited in scope”¹.

The decision clarifies that, under the UK Copyright, Designs and Patents Act 1988, AI models that do not retain training data are not themselves infringing copies of the works they are trained on, and the case maps trademark frameworks onto AI-generated content. But the question of whether the training process itself could constitute copyright infringement was left for another day—and likely another jurisdiction.