Today there was a hearing in the Getty Images v Stability AI case relating to the latter's alleged use of the former's image library to train its ‘Stable Diffusion’ AI image generation software. 

Stability AI has issued two applications. Those are applications (1) to strike out the claim and/or grant summary judgment against Getty Images, and (2) for Getty Images to respond more fully to requests for clarification about its case. Getty Images has issued an application to add a further claim relating to an ‘image to image’ feature, which is opposed by Stability AI. 

We have only seen Getty Images' particulars of claim and Stability AI's skeleton argument for the hearing. However, it seems that the main issues on the strike out / summary judgment claim are:

  1. Stability AI denies that any of the acts complained of were done in the UK, and says that the particulars of claim either do not set out a case that they were or that any such claim is “hopeless”. It has adduced witness evidence to that effect. If correct, then the proper place to bring the claim would appear to be the USA and possibly Germany (as indeed Getty Images is also doing).  
  2. Stable Diffusion does not contain the images said to have been used to train it, and in any event the model is not an ‘article’ within the meaning of the relevant legislation (s.27 of the Copyright, Designs and Patents Act 1988).  
  3. The appearance of the Getty Images watermark on some of the images  (as the associated trade marks) is not ‘use in the course of trade’ of those marks and/or is the result of the prompt used not the model itself. 

Stability AI is also seeking responses to questions such as “what text prompts were entered, by whom, and where this occurred” as well as other clarifications, and opposing Getty Images' attempt to add further infringements into the claim. 

The most interesting legal points are perhaps not the most interesting commercial ones. If the court agrees that the model was not trained in the UK, then the questions of whether such training would infringe the copyright in the underlying dataset will go unanswered. The trade mark claims have always seemed to me to be weak (getting a computer program to produce something that looks like a trade mark and infringing that trade mark being two very different things).

The hearing continues tomorrow. We will continue to watch this case with interest.