Getty Images v Stability AI Ltd [2025] EWHC 2863 (Ch)

The High Court considered whether the AI model Stable Diffusion infringed Getty Images’ intellectual property rights through the generation of images containing Getty-style watermarks and through its distribution in the UK. The court found limited trade mark infringement (in relation to certain iStock-style watermarks) but rejected most of Getty’s remaining claims.

Facts

Getty Images licenses millions of photographic and visual assets worldwide through its websites and databases, and its images commonly carry Getty Images or iStock watermarks indicating licensing requirements.

Stability AI developed Stable Diffusion, an open-source generative AI model that produces images from text prompts after being trained on large internet-scraped datasets such as LAION-5B.

Getty alleged that millions of its copyrighted images were scraped from its websites and used without permission to train the model. However, during the litigation Getty accepted there was no evidence that the training occurred in the UK, and therefore abandoned its main copyright training and development claim, as well as a related outputs claim and database rights claim.

The case proceeded mainly on three issues:

• trade mark infringement arising from synthetic images containing Getty-style watermarks,

• passing off, and

• secondary copyright infringement based on the distribution of the model in the UK.

Evidence showed that some versions of Stable Diffusion could generate images containing distorted “Getty Images” or “iStock” watermarks due to patterns learned from training data that included watermarked images.

Issues

The court had to determine:

Trade mark infringement

Whether synthetic watermarks produced by Stable Diffusion constituted use of Getty’s trade marks in the course of trade under the Trade Marks Act 1994.

Whether the signs were identical or similar and created a likelihood of confusion or other harm.

Passing off

Whether the generation of images bearing Getty-style watermarks amounted to a misrepresentation damaging Getty’s goodwill.

Secondary copyright infringement

Whether the Stable Diffusion model itself could constitute an “infringing copy” under the Copyright, Designs and Patents Act 1988 when downloaded or distributed in the UK.

Responsibility for the model’s release

Whether Stability was legally responsible for versions of Stable Diffusion released via third-party platforms such as CompVis GitHub and Hugging Face.

Judgment

The High Court (Mrs Justice Joanna Smith) largely rejected Getty’s claims, but found limited trade mark infringement.

Responsibility for releases

The court held that Stability was not responsible for the initial release of early versions of Stable Diffusion (v1.x) on CompVis platforms, meaning Getty could not rely on those releases for liability in several claims.

Trade mark infringement

Some synthetic images generated by early model versions contained watermarks resembling Getty’s marks.

However, infringement was found only in a narrow category of cases—specifically where iStock-style watermarks generated by certain model versions were identical or highly similar and likely to cause confusion.

Most other watermark examples were too distorted or contextually different to amount to infringement.

Section 10(3) claims (dilution, tarnishment, unfair advantage)

These claims failed because Getty could not show evidence of damage to reputation, dilution, or change in consumer behaviour caused by the watermarks.

Passing off

The claim added nothing beyond the trade mark issues and therefore failed.

Secondary copyright infringement

Getty’s attempt to argue that the AI model itself was an infringing copy raised novel questions, but the court rejected the claim on the facts and the statutory interpretation arguments advanced.

Implications

The decision is one of the first major UK judgments on generative AI and intellectual property, but its practical impact is limited.

The case shows the difficulty of proving IP infringement from AI outputs, particularly where outputs vary and are probabilistic.

It confirms that AI developers are not automatically liable for all outputs generated by users, especially where the system is open-source or user-controlled.

At the same time, the court accepted that AI-generated outputs can constitute trade mark use in some circumstances.

The judgment highlights that future AI-copyright disputes may hinge on where training occurs and how models are distributed, rather than the mere fact that training data contained copyrighted material.