This post has been contributed by Dr Luke McDonagh, Module Convenor for Intellectual Property.
Introduction
The relationship between artificial intelligence and copyright law is one of the defining legal questions of our time. In November 2025, the UK High Court handed down its much-anticipated judgment in Getty Images (US) Inc & Ors v Stability AI Limited [2025] EWHC 2863 (Ch) – the first major UK ruling to examine how copyright and trade mark law apply to the development and deployment of a generative AI model. The decision generated headlines. But what did it actually decide? And, perhaps more importantly, what did it leave unanswered?
Getty’s claim
Getty’s original claim was broad and ambitious. Stability AI, the company behind the Stable Diffusion image-generation model, had allegedly scraped millions of Getty’s copyright-protected images to train its model, without authorisation and without payment. Getty alleged copyright infringement, trade mark infringement and passing off. It was a comprehensive assault on almost every aspect of Stability’s conduct: the training process, the model itself, and the images it generated.
By the time the case reached judgment, the claim had been narrowed by the court, who deemed certain issues, including primary copyright infringement, out of jurisdiction. Two issues remained: (i) secondary copyright infringement and (ii) trade mark infringement arising from watermark-like features appearing in certain AI-generated outputs. The transformation from broad claim to narrow dispute is itself instructive – the reasons for this narrowing of claims go to the heart of the legal challenge posed by AI.
The Jurisdiction Problem: Where Did the Infringement Happen?
Copyright law is territorial. Broadly, infringement must occur within the UK for a UK court to have jurisdiction over it. However, the Stable Diffusion model was trained using computing resources outside the United Kingdom. This single fact proved fatal to Getty’s primary copyright claims, despite Stability’s status as a UK-registered company and the presence of UK-based developers.
Shortly before closing submissions, Getty abandoned this copyright claim relating to the training and development of Stable Diffusion, as well as all claims relating to its outputs. The central question – whether training an AI model on copyright works without a licence is itself an infringement – was never adjudicated.
This is a remarkable outcome. The case that was supposed to define the boundaries of AI and copyright in the UK ended without any finding on the issue that arguably matters most to rights holders and developers alike. In so doing, it becomes apparent that AI companies can train their models outside the UK, and UK copyright law will not reach them.
Secondary Infringement: An Intangible Article, But Not an Infringing Copy
What remained was Getty’s secondary infringement claim. The argument was creative: even if training occurred abroad, Stability had made the Stable Diffusion model weights available for download in the UK. Getty argued that Stability had imported and dealt with an ‘article’ that was an “infringing copy” of Getty’s copyright works under sections 22, 23 and 27 of the Copyright, Designs and Patents Act 1988.
Judge Joanna Smith DBE made a notable finding on the first of the two questions this raised: an intangible object can constitute an ‘article’ under the CDPA. The concept is not limited to physical things. In a world of cloud storage and downloadable model weights, that is a sensible and forward-looking conclusion.
But the claim fell at the second hurdle. The court held that for an article to be an ‘infringing copy’ it must at some point have contained or stored a copy of the copyright work in question. Both parties’ experts agreed on the underlying technical reality: the Stable Diffusion model weights do not store or reproduce any Getty images. They encode learned statistical parameters, not entire pictures. On that basis, the secondary infringement claim was dismissed. The decision is being appealed.
Trade Mark Infringement: A Narrow Finding
he one area where Getty achieved some success was trade mark infringement. In certain AI-generated outputs, Stable Diffusion had produced images bearing watermark-like features resembling the Getty mark. Getty argued this amounted to infringement under sections 10(1), 10(2) and 10(3) of the Trade Marks Act 1994.
The court’s findings here were limited. The judge dismissed all trade mark claims against Stable Diffusion XL and version 1.6, because Getty’s evidence relied entirely on outputs generated through its own litigation experiments – using verbatim or reworded Getty image captions as prompts – with no evidence that any real user in the UK had ever generated a watermarked image in that way. The court found no evidence that ordinary users would reach for a Getty caption as a text prompt.
For earlier versions 1 and 2, however, the judge concluded that at least one UK user would, on the balance of probabilities, have generated an image containing a Getty watermark. Examining specific examples, she identified two outputs from version 1 that infringed under sections 10(1) or 10(2), and one output from version 2 that infringed the Getty mark under section 10(2). These findings were described by the judge herself as ‘limited’. The broader section 10(3) claim – which would have required evidence of harm to the marks’ reputation or a change in consumer behaviour – was dismissed for lack of evidence.
Importantly, the court held that responsibility for these infringements lay with Stability, not with the users who had entered the prompts. Stability selected the training data, built and operated the system, and determined its architecture. The user’s prompt merely triggered internal processes that Stability had created.
What the Case Leaves Open
The Getty judgment is significant, but perhaps more for what it could not decide than for what it did. The core question – is it an infringement of copyright to train an AI model on protected works without a licence? – remains entirely open in UK law. Until a case is brought where training demonstrably occurred in the UK, or until Parliament acts to clarify the position, that question will go unanswered.
For rights holders, the judgment is a sobering reminder that existing legal frameworks were not designed with generative AI in mind. For AI developers, it is not the clean bill of health it might appear: the appeal is pending, the trade mark findings confirm that outputs can generate real liability, and the legislative landscape is shifting. The UK Government is consulting on copyright and AI reform. Whatever the Court of Appeal decides, this story is far from over.