Getty Images v Stability AI arguments on infringement of copyright to be appealed

Predictably as this judgment involved novel points of law, we’ve not heard the last of the Getty Images v Stability AI battle in the UK courts over infringement of copyright and the training of AI models. The Court of Appeal will now consider certain points of law to do with whether an AI model is an “infringing copy” within the meaning of the Copyright, Designs and Patents Act 1988. This case has far-reaching consequences for AI models and other intangible articles such as software so justified an appeal.

Stable Diffusion

To understand the issues to be reviewed by the Court of Appeal, it’s important to understand the way this generative AI model works. 

The relevant version of the model (known as Stable Diffusion) creates synthetic images in response to prompts entered by users. The images produced may be near identical to the images the model is trained on (including the relevant copyright works), but the works are not themselves stored in the model, nor does the model reproduce them. Instead, the weights within the model tell the algorithm how to respond to the prompts in order to produce images which are close to what the user is looking for. 

Getty argued this link - that the weights in the AI model resulted from the training data - was sufficient to constitute secondary infringement of copyright. The High Court decided against Getty Images on this point. It said that an AI model (or other intangible article) must store an actual copy of a work within itself to be an “infringing copy” within the meaning of the legislation. 

What the High Court decision means

The general consensus is that the High Court decision was a good result for AI developers and a disappointment for rights holders. It means that importing a pre-trained GenAI model into the UK is not an infringement of copyright where:

• the model was trained outside the UK; and
• does not store copies of the copyright works in its model weights.

This could be a disincentive to train and develop AI in the UK which runs counter to the government’s ambition for the UK to be a global hub for AI development. However, with permission to appeal having been granted and the ongoing review by the government of copyright and AI, it’s a case of watch this space.

The key points on appeal

The arguments on appeal around what makes an “infringing copy” will likely break down into two key questions:

1. does an infringing copy have to be an actual copy?
2. was the court correct to conclude that Stable Diffusion was not a copy of/did not contain copies of Getty’s copyright works?

On the second point, although not binding on the UK courts (because it’s a decision of the Chamber of the Regional Court of Munich), the GEMA v OpenAI case could be an interesting one for the UK courts to consider. 

The Munich Court ruled that OpenAI had infringed copyright by training its models with copyrighted song lyrics and reproducing them at the request of users. Both the reproduction of the song lyrics in the chatbot output and the technical storage and reproduction of the lyrics in the AI model were held to constitute infringement. So currently there's a divergence of opinion between the UK and German positions on copyright infringement and the training of AI models.

Incidentally, Getty also tried to appeal certain aspects of the High Court decision on trade mark grounds, however permission to appeal was not given. In our view, the trade mark points may have been moot anyway as Getty’s arguments on trade mark infringement were based on the production of images bearing the Getty Images watermark. More recent versions of Stable Diffusion include technical measures which screen images and reduce the production of images bearing watermarks so these arguments may have been of historical importance anyway. 

Another day in court

As well as the appeal in the UK on copyright grounds, Getty Images continues to pursue actions against Getty Images in the USA. The parallel US proceedings have recently been re-filed in the Northern District of California after Getty voluntarily dismissed the original action it had brought in Delaware. As Getty alleges that the training of the model took place in the US and the US case is at pre-trial stage and as OpenAI has indicated that it intends to appeal the Munich decision, headlines around AI models, training data and copyright will dominate headlines in the legal press in Europe, America and the UK in 2026 and perhaps beyond.