AI Regulation Tracker  /  Court ruling

UK court clears Stability AI on secondary copyright but finds trademark infringement in AI outputs

The England and Wales High Court held that Stable Diffusion's model weights are not "infringing copies," yet found trademark infringement where AI outputs reproduced Getty watermarks. Getty has been granted permission to appeal the copyright dismissal.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

The England and Wales High Court has drawn a sharp line between how an AI image model is trained and what it produces. In Getty Images (US) Inc and others v Stability AI Ltd, judgment handed down on November 4, 2025, Mrs Justice Joanna Smith dismissed Getty's central copyright claim against Stability AI, the maker of Stable Diffusion, while finding that the company had infringed Getty's trademarks where generated images reproduced Getty watermarks. For any business generating images with AI for UK use, the practical takeaway sits on the output side, not the training side.

What the court decided on copyright

Getty ran a secondary copyright infringement claim under the Copyright, Designs and Patents Act, arguing that the Stable Diffusion model itself was an "infringing copy" that Stability had imported into and distributed within the UK. The court rejected that theory. It held that the model weights do not store or contain reproductions of the works they were trained on, so the model is not an infringing copy for the purposes of the Act. Notably, the court accepted that an "article" capable of being an infringing copy can be intangible, but concluded the models did not meet the infringing-copy test on the facts. The result is that Getty's principal copyright case failed at first instance.

What the court decided on trademarks

The trademark claim went the other way in part. The court found infringement under sections 10(1) and 10(2) of the Trade Marks Act 1994 where some images produced by earlier versions of Stable Diffusion carried the Getty Images watermark. In its public statement, Getty framed the significance around control, saying responsibility for the presence of such marks lies with the model provider, which controls the images used to train the model. The court dismissed the separate section 10(3) claim, which requires proof of harm to a mark's reputation, on the basis that the evidence did not establish the requisite damage or change in consumer behaviour. So the trademark win was real but bounded. It rested on the specific fact that outputs carried the Getty watermark, and on the finding that the model provider, rather than the end user, controlled the training images that caused those marks to appear.

Why the output side is the compliance point

For UK-facing businesses, the duty hook is straightforward. Even where a model is trained lawfully, an individual output that reproduces a protected mark or watermark can still infringe. That liability does not depend on the copyright question at all. A marketer, agency, or publisher that circulates an AI-generated image carrying a recognizable brand watermark cannot rely on the training being lawful as a defence. Output screening, once a nice-to-have, reads after this judgment as a compliance duty for anyone publishing generated visuals into the UK market. The practical exposure is broadest for high-volume users, agencies producing client work, and publishers running generated imagery at scale, where a single unreviewed output can carry a competitor's or a stock library's mark into a public campaign. Contractual protection from a model vendor may help allocate risk, but it does not remove the primary liability that attaches to the party doing the publishing.

What the ruling does not do

The decision does not license unrestricted training on copyrighted works, and it does not settle UK copyright law for AI. Much of Getty's copyright case narrowed before trial for evidential and jurisdictional reasons, so the court decided the questions in front of it rather than the broad "is training lawful" issue that many had hoped it would resolve. It also does not immunize model providers or users from trademark exposure. Businesses should not read the copyright dismissal as a general green light.

The appeal keeps the question open

The first-instance result is not the last word. At a consequentials hearing in December 2025, Mrs Justice Joanna Smith granted Getty permission to appeal the dismissal of its secondary copyright infringement claim, declining to filter the grounds at that stage. Reports indicate Stability was refused permission to appeal the trademark findings against it. That leaves the secondary-infringement question live into 2026 and beyond, and a Court of Appeal ruling could reshape how model weights are treated under UK copyright law.

For a US reader, the case is persuasive rather than binding, but it matters. It is the first substantive UK trial on copyright and AI model training, and it sets a de facto reference point for firms operating in the UK. The immediate lesson travels well regardless of jurisdiction: manage the output, not just the model.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.