A UK court found narrow trade mark use but no infringing AI copy

The process wall

Getty Images sued Stability AI in the England and Wales High Court over Stable Diffusion, and by trial had narrowed its case: it dropped its claim that training the model infringed Getty's copyright, because the training took place on servers outside the UK, and it abandoned its database-rights and output claims. What went to judgment on 4 November 2025, Mrs Justice Joanna Smith's judgment records, was narrower: whether Stable Diffusion's model weights are themselves an unlawful copy under UK copyright law, and whether specific outputs bearing Getty's and iStock's watermarks infringed its trade marks.

What the documents show

On trade marks, the court found infringement only in named, historic instances: images generated with certain early model versions, accessed through Stability's own DreamStudio and Developer Platform tools, that reproduced an iStock or Getty Images watermark closely enough to confuse a consumer. The judgment ties each finding to a single example image and version; it does not find how often such watermarks appeared across real use. Getty's broader trade mark claim under section 10(3) failed outright, and passing off was not separately addressed. Getty's copyright claim over the model itself failed on the facts: the court held that model weights can, as a matter of statutory interpretation, count as an intangible article capable of being an infringing copy, but that Stable Diffusion in its finished form does not store or reproduce the copyrighted images it trained on, so it is not one. An earlier ruling in the same case, from 14 January 2025, confirms the parties and judge and had already narrowed Getty's representative claim before trial.

Authorship and limitation

The mechanism the judgment isolates is storage, not resemblance or training use. A watermark surviving into an output is trade mark use in the course of trade, a different question from whether the artwork was copied; a model that has learned statistical patterns from an image, without retaining that image, escapes the copyright question entirely on this court's reasoning, whatever the training process involved. Stability was also found not liable for the earliest public release of the model, since a separate research group, not Stability, published those files.

What survives

What survives is a first-instance ruling: a narrow, version-specific trade mark loss for Stability and a clear copyright win, with no finding on the scale of infringement and no additional damages awarded. Editorially, a maker should treat a visible watermark in a generated image as the strongest, most legally tested signal that training data leaked through, stronger than any claim about the model's copyright status as such.

• Read the judgment's own list of which model versions and access routes carried a trade mark finding.
• Check whether either party has sought permission to appeal, since this is a first-instance decision.
• Compare the court's reading of 'infringing copy' against the equivalent question under a different jurisdiction's copyright statute.

The ruling turns on two technical distinctions, storage against pattern-learning and a specific watermark against a general resemblance, and a maker who wants to know what the case decided should look for those distinctions before reading any headline about who won.