Getty Images UK claims against Stability AI to proceed to full trial

This article first appeared in the Intellectual Property Forum journal, published by Intellectual Property Society of Australia and New Zealand. Toby Headdon and Kyrana Hulstein, “Current Developments – Europe: Getty Images UK claims against Stability AI to proceed to full trial” (2024) Intellectual Property Forum [135].

Getty Images (US) Inc & Ors v Stability AI Ltd [2023] EWHC 3090 (Ch)

On 3 December 2023, an application for summary judgment required the UK High Court to consider the question of jurisdiction in the context of AI. After examining the parties’ evidence on this point, the judge, Smith J, dismissed the defendant’s application, citing “unanswered questions and inconsistencies” in the defendant’s evidence.

The Background

Earlier this year, Getty Images (“Getty”), a well-known provider of stock images, initiated proceedings against Stability AI, creator of the deep learning AI model and automatic image generator “Stable Diffusion”, for copyright infringement, database right infringement, trade mark infringement and passing off. In short, Getty alleges that Stability AI scraped millions of images from Getty’s websites, without their consent, and used those images to train and develop Stable Diffusion. Additionally, Getty maintains that the images created by Stable Diffusion (i.e., the outputs) reproduce its copyright works and/or trade marks. (Getty has filed parallel proceedings against Stability AI in the Delaware federal court and Stability AI is also the subject of other pending lawsuits in the US).

Stability AI has yet to file its defence to Getty’s claim. Instead, it sought to strike out various aspects of Getty’s claim, namely:

• copyright and database right infringement claims relating to works being downloaded on servers/ computers in the UK during the development and training of Stable Diffusion (the “Training and Development Claim”);
• secondary copyright infringement claims as a result of the importation of pre-trained Stable Diffusion software into the UK (the “Secondary Infringement Claim”); and
• trade mark infringement and passing off claims in relation to the synthetic image outputs generated by Stable Diffusion (the “Trade Mark and Passing Off Claims”).

Stable Diffusion had also sought further information from Getty about the particular copyright works they allege to have been infringed, for which no particulars had been provided. This request was resolved during the course of the hearing. Following clarifications to Getty’s case, Stability AI agreed not to pursue its application for summary judgment in relation to the Trade Mark and Passing Off Claims.

As a result, the Court was left to consider whether Getty had a real prospect of succeeding on the Training and Development Claim and the Secondary Infringement Claim.

Principles of Summary Judgment and Strike Out

Under Civil Procedure Rule 24.3(a), the court may grant summary judgment on the whole or part of a claim or an issue if it considers that, as noted above, the party has no real prospect of succeeding on it. In this respect:

• The court considers whether the claimant has a realistic, rather than fanciful, prospect of success (i.e., the claim is more than merely arguable). The overall burden of proof rests on the party applying for summary judgment (in this case Stability AI).
• The court must not conduct a “mini-trial”. In reaching its conclusion, the court must take into account the evidence placed before it as well as any evidence which can reasonably be expected to be available at trial.
• The court should hesitate to make a final decision in the absence of a full factual investigation, which can only take place at trial, even if the case does not seem complicated. That said, if a case is likely to fail on a point of law or construction, and the court has the necessary evidence before it, it should consider granting summary judgment.

Under Civil Procedure Rule 3.4(2)(a), the court may strike out a statement of case (such as the Particulars of Claim) or part of it on the grounds that it discloses no reasonable grounds for bringing the claim. The court is required to assume that the facts as pleaded in the statement of case are true and then ask itself whether the claim advanced on those facts has any real prospect of success. As with summary judgment, the court must not conduct a mini- trial. The court will typically accept what the claimant states at face-value unless its factual assertions are demonstrably unsupportable. Importantly, it is not usually appropriate to strike out a claim on assumed facts in a developing area of law (which is clearly applicable in this case).

The Training and Development Claim

Before considering the evidence, Smith J reiterated that “copyright (like database right) is a territorial right which confers protection on its holder only within the territory of the United Kingdom”. The key question before the Court was therefore whether there was any prima facie evidence that the training and development of Stable Diffusion took place in the UK, thereby conferring jurisdiction on the UK court.

Getty’s case for infringement was based on an inference that its copyright works were downloaded on servers and/or computers in the UK. Stability AI refuted this in two main ways:

• First, it noted that Getty had alleged in parallel proceedings in Delaware that Stability AI maintained cloud computing and physical servers in the US, trained Stable Diffusion in the US, and therefore engaged in infringing conduct in the US. Stability AI argued that this position contradicted Getty’s resistance to its application for summary judgment here, which was based on the training of Stable Diffusion taking place in the UK. Smith J disagreed; she reasoned that it was not inconsistent to allege that training occurred in the US and that, during this process, copyright works were also downloaded on servers and/or computers in the UK.
• Secondly, Stability AI contended that its evidence established that all computing resources used for the training of Stable Diffusion were located in two US datacentres operated by Amazon Web Services (“AWS”), that any services provided to it by AWS in the UK did not relate to the training and development of Stable Diffusion, and that none of its UK-based employees had ever worked on the development or training of Stable Diffusion.

Smith J considered that, had this been the trial, this evidence “would (on its face) provide strong support for a finding that, on the balance of probabilities, no development or training of Stable Diffusion” had taken place in the UK. That said, she was not satisfied that Getty had no real prospect of being able to refute this evidence or that there were no reasonable grounds for believing that a fuller investigation into the facts (following disclosure) would add to or alter the evidence available on where the training and development occurred. In particular, she identified evidence pointing away from Stability AI’s proposed factual conclusion, as well as several “unanswered questions and inconsistencies”. This supported the proposition that the Training and Development Claim had a real prospect of success and should therefore be allowed to proceed to trial.

In reaching this conclusion, Smith J, referred to: (i) human resources, and (ii) computer resources. Her analysis is underpinned by a general distrust of the evidence given by Mr Mostaque, Stability AI’s founder and CEO, which she categorised as “either inaccurate or incomplete”.

(i) Human resources

Mr Mostaque gave evidence that none of the UK-based employees in the company’s development team had worked on the training and development of Stable Diffusion. However, screenshots from interviews with Mr Mostaque suggested that individuals worked on Stable Diffusion in the UK and further evidence suggested that Mr Mostaque, who resides in the UK, was personally involved in developing the AI model. Additionally, Smith J was not convinced by Mr Mostaque’s evidence that he was confident that none of his UK-based employees had worked on the development and training of Stable Diffusion. Indeed, he had not contacted any of the employees directly (only via his HR team or lawyers), did not confirm that responses were received from each employee, nor how  exactly  they  responded. Smith J further stressed that none of the evidence had sought to explain what the training and development of Stable Diffusion had actually involved.

(ii) Computer resources

Getty’s expert witness noted that, although he could not definitively say where Stable Diffusion was developed, it may have been developed using AWS services, which may have included using and evaluating data in the UK and moving data into the UK on desktops or via other cloud services. He added that it would have been possible to train the open- source version of the AI model on a desktop computer and therefore it was likely that a combination of local laptop/ desktop devices and AWS cloud services were used to develop the model. Additionally, Smith J remained unconvinced by Mr Mostaque’s assertion that any employees who accessed the AWS cluster to work on Stable Diffusion from local devices would have done so from outside the UK. Between that and unexplained evidence of data transfers to London from AWS in the US, Smith J found there was a prospect, which was more than merely fanciful, that local devices were used during the development process in the UK.

Bearing all this in mind, and considering the conflicts of fact identified by Getty, Smith J concluded that Getty had a real prospect of success on the Training and Development Claim at trial and therefore rejected Stability AI’s application for summary judgment on the Training and Development Claim.

The Secondary Infringement Claim

Stability AI accepted that its arguments in relation to this claim would stand or fall on one point of law; namely, whether the meaning of the word “article” in sections 22, 23 and 27 of the Copyright, Designs and Patents Act 1988 (UK) encompasses only tangible items or extends to intangible items such as when software is made available on a website. Stability AI argued that it only covered tangible items and so Getty had no real prospect of succeeding in its Secondary Infringement Claim, which concerned pre-trained Stable Diffusion software.

Smith J declined to adopt Stability AI’s suggested approach for three main reasons. Firstly, she was not satisfied that Getty was bound to fail on this point, considering the relevant case law and statutory context. Secondly, she decided that the issue raised a novel question, which would be resolved once all the facts had been ascertained at trial. Indeed, this point had not previously been argued in any detail or decided by the court, especially in the novel context of secondary infringement involving internet downloading. Thirdly, there was a policy consideration as to whether an “article” should include both tangible and intangible items. For example, on Stability AI’s view, there would be an infringing copy and secondary infringement if a copy of Stable Diffusion had been imported into the UK on a memory stick, but not if it were accessed via a cloud-based service. Thus, Stability AI’s application for summary judgment in relation to the Secondary Infringement Claim was also dismissed.

Concluding Remarks

Getty’s claims against Stability AI will now proceed to a full trial (unless the parties resolve the dispute beforehand). Given the subject matter, this case is being monitored closely by lawyers as well as so-called Big Tech, other text-to-image machine learning models (such as DALL-E, Midjourney and Imagen), and start-ups built around generative AI.

As is the case with many other countries, there is presently some uncertainty in the UK regarding the extent to which using scraped data to train machine learning models infringes copyright and database rights. The UK also has a very narrow set of copyright exceptions in comparison to other countries, so one can see why Getty launched an action here.

At the end of 2022, the UK Intellectual Property Office (“UKIPO”) sought to introduce a very broad exception to permit copying for the purpose of text and data mining generally. However, this proposal was met with trenchant criticism from the creative sectors and, ultimately, was dropped. Instead, the UKIPO has convened a series of meetings with interested stakeholders to try and establish an agreed “Code of Practice” built around a licensing solution. Anecdotally, these discussions have not been particularly fruitful thus far.