Facts
- The claimants, Getty Images Inc and related companies, claim that the defendant, Stability AI, “scraped” millions of images from Getty’s websites without its consent, and used those images unlawfully as input to train and develop its deep-learning AI model Stable Diffusion and that Stable Diffusion’s output, accessed by users in the UK via a variety of means including through Stability AI’s commercial platform operating under the name of Dream Studio, was also itself infringing in that it reproduced a substantial number of Getty’s copyright works or bore Getty’s trade marks.
- Getty claims copyright infringement, database right infringement, trade mark infringement and passing off.
- The defendant has not yet filed a defence to the claim but has instead made an application seeking reverse summary judgment and/or strike out of various issues arising out of the claims before the High Court.
Issues on which summary judgment was sought
Stability AI applied for reverse summary judgment and/or strike out in respect of the following two claims in the case:
- A claim of copyright and database right infringement arising from Getty’s allegation that during the development and training of Stable Diffusion, visual assets and associated metadata, including Getty’s copyright works, were downloaded on servers and/or computers in the UK.
- A claim of secondary infringement of copyright by reason of the importation into the UK, or possession in the course of a business, sale or letting for hire in the UK, of an “article”, namely the pre-trained Stable Diffusion software, so as to infringe section 22 or 23 of the Copyright, Designs and Patents Act 1988 (CDPA).
The argument with regard to the first claim was factual in nature. Copyright, like database right, is a territorial right which confers protection on its holder only within the territory of the United Kingdom. Stability AI’s key argument was that the training and development of Stable Diffusion did not take place in the UK. It asserted that all of the computing resources it used for training were, at all times, located outside the UK and that, in particular, all of the training computing infrastructure was located in two U.S. data centres operated by Amazon Web Services Inc.
The argument with regard to the second claim hung on a point of law, namely the interpretation of the word “article” in the CDPA – in particular, whether sections 22 and 23 of the CDPA are limited to dealings in articles which are tangible things or whether they may also encompass dealings in intangible things (such as making available software on a website).
Application to amend particulars of claim to include image-to-image claim
Getty also sought permission to amend the particulars of the claim (PoC) by adding a new claim, which asserted that an image-to-image feature of Stable Diffusion, among other things, amounted to a communication to the public or had the effect of authorising acts of reproduction, so as to infringe copyright and database right. The proposed amendment introduced claims of copyright infringement, database right infringement, trade mark infringement and passing off, in relation to the image-to-image feature. Stability AI opposed this application.
The decison
The judge’s reasoning is set out in detail below. In short, she:
- Refused to strike out the training and development and secondary infringement claims, finding that both had a real prospect of success. These claims will now proceed to trial.
- Permitted the amendment of the PoC to include the image-to-image claim.
(a) Training and development claim (paras 43 to 78 of the judgment)
The question at the heart of the summary judgment application was whether there was any prima facie evidence that the training and development of Stable Diffusion took place in the UK (referred to as the “location issue”).
There was only one paragraph in the PoC which sought to connect the training and development claim with any activity taking place within the UK. It did so not by making any positive assertion of fact, but by inviting an inference that during the development and training of Stable Diffusion, Getty’s copyright works were downloaded onto servers and/or computers in the UK.
Having examined all the evidence before the court, the judge said that she was not sufficiently satisfied to be able to grant summary judgment, and that the training and development claim was manifestly unsustainable and had no real prospect of success at trial. There were also reasonable grounds for believing that a fuller investigation into the facts (for example, through disclosure and cross-examination) would affect the outcome of the location issue.
In this context, the judge observed that Getty did not hold, or have access to, the relevant information on the key question of whether any development and training may have taken place in the UK and so could not reasonably be expected at this stage to plead anything other than a purely inferential case.
Key reasons for concluding that the claim had a reasonable prospect of success stemmed from the central evidence in the arguments between the parties as to human and computer resources involved in developing and training Stable Diffusion. Stability AI’s CEO asserted that no employees worked on the training and development of Stable Diffusion from the UK. However, there were contemporaneous documents which appeared to be inconsistent with this, suggesting that the evidence would merit further investigation. Further, due to a lack of clarity regarding workers who may have developed the model, the judge held that there appeared to be a prospect, which was more than merely fanciful, that machines located in the UK could have been used during the development stage.
(b) Secondary infringement claim (paras 79 to 95 of the judgment)
Getty alleged that the following constituted secondary copyright infringement: (ii) importing into the UK an article, namely Stable Diffusion, which was and which the defendant knew or had reason to believe was, an infringing copy of the copyright works (CDPA section 22); and/or (ii) possessing in the course of a business, selling or letting for hire, or offering or exposing for sale or hire, an article, namely Stable Diffusion, which was and which the defendant knew or had reason to believe was, an infringing copy of the copyright works (CDPA section 23). According to section 27 of the CDPA, an article is an infringing copy if the making of the copy “constituted an infringement of the copyright in the work in question”.
Stability AI argued that the word “article” in sections 22 and 23 refers exclusively to tangible items and therefore does not apply to intangible software. The judge rejected Stability AI’s argument, concluding that there were good reasons to postpone to trial a decision on the correct statutory interpretation of the word “article”. This included, among others, the fact that the argument appeared to raise a somewhat novel question of interpretation of the CDPA, which would be better resolved once all the facts had been ascertained at trial because of the specific context in which the question arose (i.e., internet downloading). The judge also took note of Getty’s untested argument that there were good policy reasons why “article” should be read as meaning either a tangible or intangible article, including that, on Stability AI’s construction, “there would be an infringing copy and secondary infringement if it had brought a copy of Stable Diffusion into the UK on a memory stick but because of developments in technology which mean this can now be achieved via a cloud-based service, there would be no infringement”.
Surprisingly, no arguments seem to have been presented to the court with respect to the allegation that a trained software could qualify as a “copy” of a copyright work under section 27 of the CDPA, despite the fact that AI software generally do not contain or embed libraries of their training data.
(c) Application to amend PoC to include image-to-image claim (paras 96 to 115 of the judgment)
The claimant’s amendment application sought to introduce a new case in respect of an image-to-image feature which enables Stable Diffusion to generate a synthetic image output from an image uploaded by a user (either with or without a text prompt). The user is able to determine how closely the synthetic image output matches the image prompt by the use of an “image strength” slider, with the maximum image strength, or values approaching the maximum image strength, providing images which comprise the whole, or a substantial part, of the image prompt.
The judge permitted the amendment application. She considered that the image-to-image claim had a real prospect of success and that it was plainly a claim that should go to trial
Comment – significance of the decision
While the decision only dealt with the limited question of whether the selected claims had a reasonable prospect of success, which is not a very high bar to satisfy in such a complex case, it offers an insight into some key factors that will be relevant to developers and providers of AI models.
- The first is that for the purposes of any claim of infringement relating to the input phase of training and developing an AI model, the key question will be whether the training and/or development took place in the UK. If not, there will be no infringement under English law. The judgment also offers a glimpse into the kind of evidence that the court may look at to establish whether research and/or development (to the extent this involves different processes or people) took place in the jurisdiction. This includes checking (i) the location of the development teams (including UK and non-UK employees as well as independent contractors), and (ii) the resources used for the development, with the key questions likely being what machines were used for the development and whether there was any kind of data transfer to the UK. Determination of whether training and/or development of an AI model took place in the UK will require a granular examination. In order to avoid liability, developers of AI models should ensure that no part of the training or development is carried out in the UK (and, preferably, that there is evidence supporting this).
- The second is that the judgment left open the possibility that an AI model could qualify as an infringing copy of its training data, thereby suggesting that AI models could be embedding a copy of their training data, and also held that secondary infringement of importing, possessing or dealing with an infringing copy, a type of infringement traditionally reserved to tangible objects, may be able to cover intangible things such as making available software through a website, thereby signaling the Court’s openness to a potential departure from classic copyright and information law principles. We view the chances of success of such argument as low, for reasons which do not seem to have emerged during the summary judgment discussion, including the implausible compression of billions of training data points within the Stability AI model.1
Notwithstanding the above, the judgment makes clear that to succeed at trial, claimants will have to demonstrate clearly how the AI system in question works if they are to prove that an infringement has taken place in the UK.
- A fact that has already been discussed in the context of other AI-related cases, notably the case opposing Sarah Silverman to Stability AI and others, which is pending before the U.S. District Court for the Northern District of California.
In-depth 2024-044
