High Court rules that Getty v Stability AI case can proceed

While most of the attention regarding artificial intelligence and copyright has been centred in the United States, it’s useful to remember that one of the main pieces of litigation is taking place in the courts of England and Wales. We’ve now had the very first order emanating from the case, this is Getty Images v Stability AI [2023] EWHC 3090 (Ch) (copy here).

The case was filed earlier in the year, so this order deals mostly with preliminary issues, as well as giving an insight into some of the preliminary evidence presented by the parties in a hearing in October. For those unfamiliar with the case, Getty Images is a well-known provider of stock photographs, and Stability AI is the UK generative AI startup, which also has offices in the US. Getty sued Stability AI for copyright infringement, database right infringement, trade mark infringement, and passing off.

The core of the claimants’ allegations is that Stability AI scraped millions of images from the Getty website without consent, and used these images to train and develop Stable Diffusion. It’s further claimed that the synthetic images generated by Stable Diffusion, accessed by users in the UK, infringe upon Getty Images’ copyrighted works and bear their trade marks​​. Some of these images had been presented in the particulars of claim, but it was never made clear how the images came to be. I was able to produce some images myself with older versions of Stable Diffusion bearing the semblance of a Getty Images logo, but none of the outputs produced appeared to come from the images in the input. The idea here is that Stable Diffusion “memorised” the Getty logo, and could place it on outputs on demand. This is no longer possible as far as I can tell.

Prompt: Old image of a cinema, Getty Images. Generated with Stable Diffusion around February 6 2023.

It’s important to stress that the Defendant hasn’t yet filed their defence to the claims, they have applied for reverse summary judgment and/or strike out of various issues arising from the claimants’ claim​. One application was a request for there to be a clear identification of which works are alleged to have been infringed, the complaint contained references, but no actual examples of alleged infringement. This is a common thread that is starting to emerge in most of the current litigation, and one that claimants will have to address going forward. It’s not only good that you claim that your works were infringed, you need to provide at least some evidence that they were. However, any samples provided will be privileged and not disclosed in the proceedings.

An application to have the claims struck was never going to succeed in my opinion, the parties argued on the well-established principle set out in Easyair Ltd v Opal Telecom Ltd [2009] EWHC 339 (Ch), and that is that to proceed the Claimants must have at least a realistic chance to succeed, and not just a fanciful one. Justice Smith unsurprisingly ruled that this was the case.

The most important aspect of the hearing was whether the training and development of Stable Diffusion occurred within the United Kingdom, a point critical to the Defendant’s application for reverse summary judgment. In short, if all of the training took place outside of the UK, the English court could rule that it has no jurisdiction over the matter, and the case can continue in the US courts, or wherever else the training took place. This is interesting for several reasons, particularly because I have been arguing for a while that jurisdiction will play a vital role in AI copyright cases. This question is vital because, as Justice Smith points out, copyright (and database right) claims are strictly national in nature, citing Starbucks (HK) Ltd v British Sky Broadcasting Group plc [2015] UKSC 31, Smith J also cites Prof Wadlow in this point: “the nature of copyright as a legal right of property arising in any given jurisdiction from national legislation, common law or self-executing treaty, means that it must be wrong to speak as if there were a single international copyright”.

So it was up to the Defendant to try to have the English leg of the case dismissed, they presented evidence suggesting that all training and development occurred outside the UK, primarily in the United States, and involved high-performance computing resources​. The evidence was that the database used had been LAION, which was collected in Germany, and the actual training itself took place at different other stages. Expert evidence was given on that fact, and Smith J found it uncontested that training of an AI would involve not only the dataset, but also includes “the development of the neural architecture of the model followed by several stages of iterative training and evaluation.” But where did that actual training and development take place?

The Defendant presented evidence that Getty was also suing in the US, and that in that complaint they had alleged that Stability AI maintained servers in the US, specifically using AWS. Emad Mostaque, the boss at Stability AI, and various employees testified that no training had taken place in the UK. I thought that this point could have been enough on its own to have the case struck down in the UK, but Justice Smith was not entirely convinced by this evidence presented by the Defendant, pointing out that it might not fully reflect the situation and that further investigation could reveal more pertinent facts​​. Given that there is doubt as to where the training took place, the case could continue in the English courts. Justice Smith comments:

“I observe that if this were the trial of this action, the evidence to which I have referred above would (on its face) provide strong support for a finding that, on the balance of probabilities, no development or training of Stable Diffusion has taken place in the United Kingdom. This is not, however, the trial of this action […], I must be satisfied that there are no reasonable grounds for believing that a fuller investigation into the facts would add to or alter the evidence available to the trial judge and so affect the outcome of the Location Issue.”

Smith J continues:

“Having examined with care all the evidence before the court, I am not so satisfied. There seems to me to be (i) evidence potentially pointing away from the factual determination on the Location Issue that I am invited to reach by the Defendant; (ii) evidence raising unanswered questions and inconsistencies relevant to that determination; and (iii) reasonable grounds for believing that disclosure may add to or alter the evidence relevant to the question of where the training and development of Stable Diffusion took place. All of this clearly supports the proposition that the Training and Development Claim has a real prospect of success and must be permitted to go to trial.”

This here is very much the core of the order. Almost, but not quite over the line. Justice Smith goes through some of the evidence presented, pointing out that there were just enough inconsistencies in some of the evidence presented by the Defendant to open the possibility that some training had indeed taken place in the UK. Smith J concluded that there is a real prospect of success for the Training and Development Claim, indicating that it should proceed to trial. This decision was influenced by evidence that potentially contradicted the Defendant’s position, raised unanswered questions, and suggested that further disclosure might change the evidence relevant to where Stable Diffusion’s training and development occurred. An important part of the doubt were a series of tweets and interviews that seemed to contradict the Defendant’s statements.

Concluding

This case is still at the very early stages. I’ve already read a couple of sensationalist press releases stating that the judge didn’t believe Stability AI. That is not what happened, the judge said that on the prevalence of evidence she would lean towards the training not having taken place in the UK, but that there were enough questions unanswered with regards to the evidence presented as to continue to trial and not strike down on location grounds. Social media strikes again.

Edited: We’re still at the very early stages, it’s so fascinating that this case could have finished right there and then had the defendant not made social media declarations. There are a few interesting aspects that could be inspected. The trial could still rest on the location question, if no training took place in the UK, then it becomes about the outputs, and I don’t think that outputs are infringing specific copyright, which is where trade marks and passing off come in. But if that is the case, who is passing off? Is it me prompting “Getty Images” on the link above?

Finally, the Defendant hasn’t issued their full defence yet, but you can sort of see where things are going. I think that the Defendant will insist of the lack of evidence of which works were infringed specifically. And maybe on the outputs. It will always rest on the outputs.

Selfishly, I just hope that no other big AI legal development takes place during my holidays, I’d hate to have to interrupt my Margaritas to read another case.