Getty Images v Stability AI: A long-awaited judgement that leaves AI stakeholders waiting longer

In a long-awaited decision, the High Court substantially dismissed Getty Images’ claim against Stability AI, finding no secondary infringement of copyright and only limited, historic trademark infringement.

For observers hoping for a landmark precedent on how existing copyright law applies to AI training, the judgement was disappointing, as Getty dropped its primary training claim for jurisdictional reasons.

However, on 16 December 2025, Getty received permission to appeal the secondary infringement aspect of the judgment, reigniting the question of what can constitute an ‘infringing copy’.

Background

In January 2023, Getty Images, a well-known stock photo agency, sued Stability AI, the company behind Stable Diffusion, a generative AI model.

Getty primarily alleged that Stability had used millions of Getty’s images without permission to train and develop Stable Diffusion, amounting to primary copyright infringement. Getty also claimed that the models’ outputs infringed its copyright and trademarks, and brought related database rights and passing off claims. However, by the time of the full trial in June 2025, the case had considerably narrowed.

Getty abandoned its primary copyright infringement claim, accepting that it could not prove the training took place in the UK. This also removed the database rights claim. The court was therefore left to consider secondary copyright infringement, trademark infringement, and passing off.

In a 219-page judgment, the judge rejected the secondary copyright infringement claim in full, found only “historic and extremely limited” trademark infringement, and declined to address passing off because neither party made detailed submissions on it.

Getty have since been granted permission to appeal the unsuccessful secondary infringement ruling.

The focus of this article is why the secondary copyright infringement claim failed and why Getty are appealing it.

How a generative AI model like Stable Diffusion works

The court questioned how generative AI models like Stable Diffusion operate and a brief explanation is important to understand the court’s reasoning regarding secondary infringement.

Training

The model is trained on an extensive data set of images so that it can learn to associate text with patterns such as shapes, colours, and textures. It does this while simultaneously learning how objects relate with each other.

It does not store the images it processes. Instead, it repeatedly adjusts billions of numerical ‘model weights’ that represent what the model has learned from analysing the images.

Generation

When a user enters a text prompt, the model begins with random visual noise and gradually reshapes it into an image. Using its learned model weights, it applies shapes, colours, and textures that it identifies as appropriate to the prompt.

By iteratively removing noise and adding structure, the model produces a new image that reflects learned patterns, rather than reproducing any specific image from its training.

Secondary infringement under the CDPA

Because of the jurisdictional difficulty with primary infringement, Getty relied on secondary infringement. Sections 22 and 23 of the Copyright, Designs and Patents Act 1988 (CDPA) provide that copyright is infringed where a person imports, possesses, or deals with an article which is, and which they know or have reason to believe is, an infringing copy of a work.

Getty argued that, even if Stable Diffusion had been trained outside the UK, importing the trained model into the UK amounted to secondary infringement. However, there was now a new obstacle. Getty had to show the AI model was both:

An ‘article’
An ‘infringing copy’

Was the AI model an ‘article’?

The term ‘article’ is not defined in the CDPA, and the judge reasoned that the word acquires its meaning from its surrounding context. Ultimately, the judge determined that an intangible AI model stored in the cloud could be an ‘article’ for the purposes of secondary infringement. Getty therefore succeeded in establishing that an AI model is within the scope of being an ‘article’.

Was the AI model an ‘infringing copy’?

The more difficult question was whether the AI model that was created by processing infringing copies of images was itself an infringing copy. The judge reasoned that it was not, instead finding that an infringing copy must itself be a copy of the work. In other words, there must be a sufficient degree of reproduction of the original work in the infringing article.

In the evidence presented, Stable Diffusion never consisted of, stored, or contained copies of Getty’s images. It contained only model weights derived from statistical analysis of those images. The model and its outputs were therefore not literal copies of Getty’s works within the meaning of section 27 of the CDPA.

Since neither Stable Diffusion nor its output images were held to be infringing copies, the secondary infringement claim failed.

Interestingly, the threshold of constituting a ‘copy’ under the secondary infringement claim was arguably trickier than what would have been required with the original primary infringement claim. This is because for primary infringement the concept of copying can be broader and encompass non-literal copying.

Appeal

The judge recognised that her interpretation of an ‘infringing copy’ was both novel and could have wide-spread implications for AI models and intangible articles generally. Appreciating that it was an issue where “the minds of reasonable lawyers may differ”, the court granted permission for Getty to appeal the secondary infringement aspect of the claim.

Key takeaways

Getty established that an intangible AI model can be an ‘article’ for the purposes of secondary copyright infringement. The decision at first instance provided that to be an ‘infringing copy’, the model (or article) must reproduce or retain a discernible copy of the protected work, rather than merely using information derived from it. However, this issue will be considered again on appeal.

Irrespective of the appeal, since Getty’s claim narrowed, there are key questions left unaddressed. Particularly whether training an AI model in the UK on copyright-protected works can amount to primary copyright infringement. AI developers and users will be closely watching Getty’s parallel proceedings in the US, as Stability maintains the US is where the model was trained, keeping the core primary infringement issue alive.

This article was co-authored by Trainee Campbell Kershaw.

Contributors: