Who is involved?

Getty Images (Getty), owner of the iStock brand, is a global visual content provider offering licensed media through its Getty and iStock websites. Since 1995, it has licensed content to users in over 200 countries.

Stability AI (Stability), founded in 2019, develops generative AI tools, including deep learning models for image, music, and text generation. Its image models, released under the Stable Diffusion name, include versions v1.x, v2.x, SD XL, and v1.6.

What are the claims?

Due to various evidential and procedural challenges, Getty’s claim narrowed significantly as proceedings progressed. Most significantly, Getty could not produce evidence that the training and development of the Stable Diffusion models took place in the UK. Stability had already blocked the prompts which Getty claimed had been used to generate the examples of infringing output. The primary copyright infringement claims and the database rights infringement claim were therefore dropped.

The key issues left to be decided following the close of evidence at trial were:

• Secondary copyright infringement: Getty argued that the Stable Diffusion v1, v2, XL, and 1.6 model weights were infringing copies. By importing and making them available in the UK, Stability allegedly breached sections 22 and 23 of the Copyright Designs and Patents Act 1988 (CDPA).
• Trade mark infringement and passing off: Getty claimed Stability violated Getty and iStock trade marks under sections 10(1), 10(2), and 10(3) of the Trade Marks Act 1994 (TMA), as some outputs included their watermarks without being prompted. Getty also alleged passing off based on the same conduct.

Summary of the judgment

Getty’s secondary copyright claim failed entirely. The judge found that “article” under the CDPA could include intangible electronic copies, but since Stable Diffusion models do not store or reproduce copyright works, they were not “infringing copies” under sections 22 and 23.

On trade mark infringement, Getty showed some user-generated images with Getty and iStock watermarks, though most examples came from its own tests. The judge was only prepared to find infringement where UK users had actually generated such images.

Getty proved that the v1.x models infringed the iStock Marks under section 10(1) TMA, but not the Getty Images Marks. It also succeeded under section 10(2) TMA for both iStock and Getty Images marks with the v1.x and 2.x models. However, there was no evidence of UK users generating watermarked images with the SD XL and v1.6 models, so no infringement was found for those.

Getty’s section 10(3) TMA claim was dismissed entirely. The judge found no evidence of dilution, reputational harm, unfair advantage, or any change in consumer behaviour.

The judge also declined to address the passing off claim in any detail. Getty had not fully argued the point, and neither party took the opportunity to make further submissions.

Key takeaways

This case provides a number of key takeaways, both for AI developers of, and for those hoping to protect their content from being used to train AI models.

Implications for AI developers:

English law has not yet confirmed that training on copyrighted data is automatically infringing – but training models on data outside the UK definitely is not considered an infringement.

Ensuring that your AI models function without reliance on storing or reproducing copyright works will prevent claims that they are “infringing copies” and mitigate the risk of copyright claims under the CPDA.

Outputs can be unpredictable, however incorporating improved prompt filtering and output moderation, as was done in Stability’s newer models, can successfully prevent the generation of infringing outputs and reduce your legal risk.

Developers are liable for outputs, even if users control the prompts. This is due to the actions taken by developers when making the models available and making deliberate choices as to the content and make-up of the dataset on which the models are trained and the filters to be applied, and writing the code that users download when accessing the models.

Implications for content owners:

AI models can be subject to IP claims, and outputs may infringe trade marks. However, some challenges in bringing legal action include:

• Jurisdictional issues: AI training often occurs across borders, complicating enforcement due to the territorial nature of copyright.
• Storage/reproduction: Many models use statistical representations, instead of storing or reproducing images. This means that no “infringing copies” have been created.
• Evidence: in spite of being a large, well-resourced company, Getty struggled to gather sufficient evidence for their claim. This was made difficult due to the complexity of the models and distribution channels, and lack of transparency requirements surrounding how AI models work and what is used to train them. Additionally, where end users generate outputs without sharing or reporting what has been generated, it is difficult to appreciate and find evidence of the scale of real-world infringing outputs generated.

How we can support you

This case underscores the ongoing uncertainty around AI and IP law, offering limited clarity. Ultimately, the case was in many ways specific to its facts and the judgment’s scope is narrow, as the judge herself noted. The key question of whether training AI on copyrighted material constitutes infringement in the UK remains unanswered. In the absence of clear legislation, courts are proceeding cautiously, suggesting that government intervention will be needed to define the legal boundaries.

As AI becomes increasingly central to business operations, these legal issues will grow in importance for developers, content owners, and users. Our expert team is ready to guide you through this still largely uncharted legal landscape.

Please contact Alan Harper or John-Joe Massey to discuss how we can support you.