Intellectual property (IP) practitioners, creatives, and AI developers have been eagerly awaiting the judgment Getty Images v Stability AI [2025] EWHC 2863 (Ch) as it is the first claim against an AI developer to be heard in English courts. Following the publication of the judgement, was it everything we had hoped for?
The claim originally served by Getty on 12 May 2023 covered both primary and secondary copyright infringement, database right infringement, registered trade mark infringement, and passing off. By the end of the trial, only the claims relating to secondary copyright infringement, registered trade mark infringement, and passing off remained. English courts are therefore still yet to grapple with primary copyright infringement, which is the direct copying of copyright works, in the context of AI. However, there are still points of interest arising out of this judgment.
Secondary copyright infringement
The key statutory provisions in this case are sections 22, 23, and 27(3) of the Copyright Designs and Patents Act 1988 (CDPA). Section 22 of the CDPA prohibits the importation into the UK of an “article” which a defendant knows or has reason to believe is an “infringing copy” of a copyright work. Section 23 is closely related, covering the possession or dealing of an “infringing copy”. An “infringing copy” includes an article that would infringe if the making of it in the UK would have constituted copyright infringement (section 27(3)).
It was alleged that Stability AI had used millions of images from Getty’s website to train Stability’s AI model Stable Diffusion or more precisely, its ‘model weights’. Model weights are the parameters that an AI model learns during training and they are involved in the process of turning inputs (e.g. an original photo or text-based prompt) into outputs (e.g. an AI-generated image). Model weights are altered by what they are exposed to during training. Critically, as acknowledged by Getty, the model weights do not store or reproduce the images they were trained with.
The key issues were:
- For the purposes of the CDPA, could an “article” be intangible such that model weights constitute “articles”?; and
- Ff the making of an article involved the use of copyright works but never actually contained/stored the copyright works, could that article constitute an “infringing copy” for the purposes of the CDPA or to put more simply, could the model weights constitute an infringing copy of copyright works even if they never contained/stored those works?
Getty was successful on the first point. The judge found that an “article” could be intangible. As part of its pleadings, Getty reminded the judge of the “always speaking” principle of statute interpretation. This means changes that have occurred since the enactment of the relevant statute should be taken into consideration such as, like in this case, developments in technology.
Getty was unsuccessful on the second point. The judge found that “an infringing copy must be a copy… I cannot see how an article can be an infringing copy if it has never consisted of/stored/contained a copy”. The judge went on to explain that “[w]hile it is true that the model weights are altered during training by exposure to Copyright Works [images owned by or exclusively licensed to Getty Images (US) inc], by the end of that process the Model itself does not store any of those Copyright Works; the model weights are not themselves an infringing copy and they do not store an infringing copy. They are purely the product of the patterns and features which they have learnt over time during the training process.” Getty’s claim of secondary copyright infringement was therefore rejected.
The “always speaking” principle could not get Getty over the line when it came to defining an “infringing article” in the modern-day context of AI training. Of course, courts cannot create law; however, the case highlights that the training of an AI model presents a unique scenario: training can involve the use of existing data, including images, in which IP rights may subsist, yet the trained AI model may not contain reproductions of the material it was trained on.
Trade mark infringement
Getty’s claims of registered trade mark infringement and passing off stem from images generated by Stable Diffusion that include Getty’s watermarks, noting that these were often distorted.
Getty faced an uphill battle proving these claims. Getty was able to demonstrate that Stability Diffusion could be pushed into producing images that included the Getty watermarks, but was unable to evidence that real-life UK users of later versions of the model had generated any images showing the watermark. Stability successfully argued that the prompts Getty used in producing the images it submitted as part of its evidence were contrived. As a result, Getty’s success was limited, and was restricted exclusively to historic versions of Stability Diffusion (as later versions of Stable Diffusion included filters to remove the Getty watermarks).
The judge declined to address passing off.
Observations
When the case was filed in January 2023, it was hoped that it would provide clear guidance on key IP issues concerning AI outputs, particularly in relation to the copyright issues. Through ten interim hearings, Stability AI successfully narrowed the scope of the claims in the UK such that the High Court decision was inevitably going to be of far less significance to those considering these legal issues, particularly on copyright. The findings in relation to registered trade mark infringement were always going to be, at best, subordinate.
More broadly, the case highlights the limitations of the CDPA (which after all, came into force on 1 August 1989) as an effective mechanism to protect the commercial interests of creatives, through the effective enforcement of IP rights which subsist in those works. In the absence of successful appeal, it would now appear that Parliament must step in if those interests are to be protected.
It is important to also keep in mind that the parties are embroiled in litigation in the US, where the Stability Diffusion model was trained. This case may therefore be more significant as it should deal with the issue of using works in which copyright subsists to train AI models without the authorisation of the copyright owner, unless of course Stability AI successfully persuades the court that it does not need to deal with this.
If you require advice and assistance on copyright and trade mark infringement cases, please contact Will Sander and Jennifer Stratfold.