UK High Court rules on Getty Images v Stability AI

Background

On 4 November 2025 Mrs Justice Joanna Smith of the High Court of England and Wales gave judgment in Getty Images v Stability AI, a case about whether the Stable Diffusion image model, in its training and operation, infringed Getty's rights.^[1] By the time the case reached judgment it was narrower than it had once been: Getty abandoned its primary infringement claims during the trial, leaving the court with what has been described as a diminished case.^[3]

The legal framework the court faced

Copyright claims of this kind usually come in two layers. Primary infringement concerns the acts of copying themselves — reproducing a work, or using it in a way the law reserves to the rights holder. Secondary infringement concerns dealing in the results: importing, possessing, selling or otherwise handling copies that are already infringing. A secondary claim therefore depends on showing that the object in question is an "infringing copy".

Applying that framework to a generative model is not straightforward. A model's weights are numbers produced by training rather than a library of stored pictures, and whether such an object can be an infringing copy at all was exactly the question the case raised.

The copyright holding

The court rejected Getty's secondary copyright claim.^[1] It accepted that an AI model can be an "article" — the threshold question of whether the object falls within the scope of the secondary-infringement provisions — but held that Stable Diffusion is not an "infringing copy", because it does not store reproductions of the works.^[1] The reasoning turns on what a model is: on the court's analysis, the model does not contain copies of the training images, so it cannot be an infringing copy of them.

The trademark findings

Getty did not leave empty-handed. The court made narrow findings of trademark infringement based on watermarks appearing in some of the model's outputs: ISTOCK watermarks in some outputs from v1.x versions, and watermarks resembling GETTY IMAGES in some outputs from v2.1.^[1] The judge described these findings as "extremely limited in scope".^[1]

Claims founded on reputation were rejected, and the passing-off claim was not decided.^[1]

Why it matters

The judgment has been described as significant for the balance to be struck between content creators and the AI industry.^[3] For creators, the outcome shows that a case built on the model itself, rather than on particular inputs or outputs, runs directly into the question of what a model actually contains. For AI developers, the limited trademark findings point to a separate exposure: watermarks that survive into generated images can still give rise to liability even where the copyright claim fails.

The dispute is not over. A CMS case tracker reports that permission to appeal the secondary-infringement point was granted in December 2025.^[2] An appeal would put the central holding — that a model which does not store reproductions is not an infringing copy — before a higher court.