UK Court Draws a Narrow Line on AI Training and Copyright

In a landmark decision issued this week, the High Court of England and Wales offered the first significant judicial look at how UK copyright and trademark law apply to AI image generators. In Getty Images v. Stability AI, the court stopped short of resolving the biggest question of whether training AI models on copyrighted material is itself infringing, but it provided important guidance for both developers and rights holders navigating the uncertain terrain of generative AI.

Getty Images had alleged that Stability AI unlawfully used millions of its copyrighted photographs to train Stable Diffusion, its popular text-to-image generator. It also claimed that the model’s outputs, some of which appeared to include Getty’s watermark, infringed its trademarks and diluted the value of its brand.

Justice Joanna Smith found that the Stable Diffusion model itself "does not store or reproduce" any copyrighted works and thus cannot be considered an “infringing copy” under the Copyright, Designs and Patents Act of 1988. The judge also accepted that the training and development of the model took place outside the UK, an important jurisdictional fact that undermined Getty’s theory of infringement.

The result is a significant, if narrow, win for AI developers. The court’s reasoning suggests that a model trained on copyrighted data outside the UK, and which does not retain identifiable copies of works, may not give rise to secondary infringement under UK law.

Getty did succeed in part on its trademark claims. The court found that some AI-generated images reproduced the Getty watermark or logo, constituting trademark infringement under sections 10(1) and 10(2) of the Trade Marks Act 1994. However, these findings were also limited. Justice Smith described her ruling as “historic but extremely limited in scope,” noting that there was insufficient evidence to show widespread misuse of Getty’s marks.

This aspect of the decision underscores a practical risk for AI companies: even if copyright claims fail, visible trademarks or watermarks appearing in model outputs can still expose developers to liability.

The decision leaves unresolved the core issue of whether training AI systems on copyrighted content without permission constitutes infringement. Because Getty withdrew much of its direct infringement claim mid-trial, the court did not address whether ingesting copyrighted works during training could itself breach UK law.

That question remains one of the most consequential in modern IP law. As governments in the UK, EU, and U.S. continue to debate text-and-data-mining exceptions and transparency requirements, today’s decision may fuel legislative rather than judicial clarification.

Key Takeaways for Legal Practitioners

• Jurisdiction is critical. Where and how model training occurs can determine the outcome. Developers should document server locations and data flows to manage territorial risk.

• Storage matters. The absence of stored or reproducible copies of training data weighed heavily in Stability AI’s favor. Internal documentation showing “no-copy retention” is a key compliance safeguard.

• Trademark hygiene is essential. Developers should filter datasets and outputs for watermarks or logos, while rights holders may find trademark law a more effective enforcement tool than copyright, though both may have better success if secondary liability is available.

• The fight is far from over. Largely consistent with rulings in the US, this ruling offers breathing room for AI companies but little comfort for rights holders. The next major test in the UK will likely involve a model trained domestically, forcing courts to grapple directly with whether “learning from” copyrighted works is copying at all.

For now, Getty v. Stability AI marks a cautious step toward legal clarity, but only a step.