The relationship between an AI model and copyright law through the lens of Getty Images v Stability AI Limited

On the 4th November 2025, a judgment by Mrs Justice Smith (DBE) was handed down where Getty Images failed in their secondary infringement copyright claim against an AI company. Despite a decision by the Court of Appeal being expected soon, the High Court decision may offer some useful insights into the entanglements between copyright law and AI technologies.

Who are the parties?

The Getty Images business was founded in 1995 and makes photographs and videos, with their associated metadata and captions available through their websites. More precisely, each visual asset contains a uniform resource locator (URL) together with a tag containing a caption for the image (para 21). These materials are acquired lawfully through licencing agreements. On the other hand, Stability is a relatively new company and was founded on 4 November 2019. It is an AI company, specialising in machine learning software. The case concerns one of their AI models known as Stable Diffusion.  

What is Stable Diffusion?

Copyright claims

Getty abandoned their copyright claim regarding the training and development of Stable Diffusion as the training appears to have taken place outside of the UK. Instead, Getty argues Stable Diffusion is infringing copy imported into the UK, as the making of the model would have constituted an infringement of copyright in the UK. Their allegation relates to what is termed in copyright as secondary infringement according to s22 and s23 of the Copyright, Designs and Patents Act 1988 (CDPA).

The intangible nature of the model weights in an AI model

For a secondary infringement claim to be successful, the importing infringing copy must be classed as an article. The article, in question, are the model weights in the AI model.  According to Mrs Justice Smith, the model weights are an intangible record of information derived from the training data (para 601). In my view, this interpretation is correct as it aligns with Lemley and Cooper’s seminal work.¹ The model weights do not retain a discrete copy of any copyright works used to train the model. Instead, information from the training materials is encoded in a way that determines how the model distributes probability across sequences of tokens that represent text (Lemley and Cooper, 2026, p. 9). For example, the model may assign a higher probability to the next token being ‘mouse’ from the training materials instead of ‘elephant’ when responding to a prompt that states ‘the cat ran after the’.² Considering Lemley and Cooper’s research, finding the model weights are an intangible article appears to have a sound basis.  

An infringing copy: the tension between tangible and intangible matter

Can an intangible article such as model weights amount to an infringing copy?

Although Mrs Justice Smith explains that an infringing copy can be an electronic copy stored in an intangible form, they conclude that the model weights are not an infringing copy, as no copyright work was stored in the model weights. In other words, Stability did not import into the UK an infringing copy (paras 597, 600). Mrs Justice Smith examined provisions beyond s22, s23, and s27 (which provides the meaning of an infringing copy) to s17 for an understanding on copying. The reasoning provided by Mrs Justice Smith reflects the plain meaning of s17(2) which states that copying is in relation to reproducing an artistic work in any material form (para 597). Considering Lemley and Cooper’s research, a plain meaning of this provision would suggest that there are no reproductions of artistic works present in the model weights. Logically, if there are no copies of a copyright work (artistic works) in the AI model weights, the imported copy cannot be an infringing copy. However, it appears disagreement arises between the parties about whether Mrs Justice Smith should have drawn on s17 to shed light on the provisions dealing with secondary infringement.  

The relationship between s17 and secondary infringement

The inclusion or exclusion of s17 (the meaning of copying) is likely to form a crucial part of the appeal. The reasons for inclusion appear to align with Sony v Ball, a case referred to by Mrs Justice Smith. In that case, Laddie L.J. refers to s17 when interpreting the meaning of an infringing copy (p.168). In sharp contrast, Getty contends that s27 on an infringing copy is sufficiently clear without the need to consult s17. If this argument is accepted, Getty purports that all that is required is that the making of the AI model in the UK would have constituted copyright infringement. In other words, there would be no need to prove that there was a reproduction of a copyright work in the model. Overall, the inclusion or exclusion of s17 is likely to be decided by the Court of Appeal, however, the case reveals more about copyright law.

Concluding remarks

Ultimately, whether an infringing copy requires a reproduction of a copyright work or not extends beyond statutory interpretation to how copyright law acts as a vehicle for control and power over information. It appears copyright law is at a crossroads and will need to decide how the law can meaningfully protect authors while recognising the importance of technological innovation.

Footnotes

[1] Lemley is a professor at Stanford Law School and partner in Lex Lumina LLP.  Cooper is an assistant professor in computer science at Yale University.

[2] Lemley and Cooper draw on different examples, p.9.

Dr Karen O’Neill

Karen is an associate lecturer for the Open University Law School. Karen has a PhD in copyright law over computer-generated works from the University of Birmingham. At the Open University, Karen teaches criminal law (W111), contract law (W212), explores modern legal issues at postgraduate level (W811) as well as supervising dissertation projects (W350).