The Getty Images v Stability AI [1] case has been highly anticipated for several years. In this article we examine the AI-driven copyright issues that arose during the case; analyse Mrs Justice Smith’s decision (including what, in our view, the court got right, and perhaps didn’t); and consider the future of copyright protection in a world increasingly dominated by AI models and their voracious appetites for large-scale training data. In short, are we dealing with our very own real-life J.A.R.V.I.S., here to wait on us hand and foot? Or, to borrow an expression from Breaking Bad’s creator Vince Gilligan, is every sophisticated model merely a “plagiarism machine”?
AI, its Model Weights, and Copyright Conundrums
Ada Lovelace’s Analytical Engine has long been surpassed by computer algorithms and enormous models (at the time of writing, most state-of-the-art large language models (LLMs) are estimated to consist of several trillion parameters) that can learn at phenomenal speeds. Moreover, despite Lovelace’s clarification that “the Analytical Engine has no pretentions whatever to originate anything”, the assumption that computers and algorithms cannot produce original work seems increasingly open to doubt. In the past three years, AI has gained the ability to produce novels, drawings or music with just a few lines of input prompts. While the actual originality of the works it produces remains to be seen, one must wonder where the ability to instantly produce a seemingly original work comes from.
The answer lies in how the AI is trained. Like a child, an AI model must learn from repeatedly being shown examples, requiring terabytes of the relevant data type or types, so that it may identify the underlying patterns. How do the words in a language typically connect together? What does a cat look like? Which notes complement an E major chord? This learning is aptly named “machine learning”. Machine learning isn’t a new concept; it’s been widespread for some time, for example in simple route-planning services in which an algorithm is consistently exposed to new data for trial-and-error learning, enabling it to predict the optimal path for one’s route to the office or theatre. What has changed is computers’ capacity for this kind of learning, which has increased exponentially with the introduction and wide-scale popularisation of neural networks. These neural networks are built to emulate the human brain and its neurons, where each of a large number of “artificial neurons” exists as a specific information-processing unit with a specific set of parameter values. These neurons are then linked, via their outputs, to the inputs of other neurons which in turn process another value, and so on, with all of the many links, known as “weights”, establishing a certain kind of assumed relationship between inputs and outputs. The magic here is in the fact these weightings can be changed during AI training enabling the system to learn like a child would. Once enough neural network layers have been stacked together, this leads to “Deep Learning”. These weights provide the reason why one model (say, ChatGPT) may give a different output from another (say, Claude) for the same prompt; the different training data provided to fine-tune the models’ weights will lead to a different output from the models.
This is where the case for Getty, and other copyright owners, begins to emerge. Stability AI’s deep learning model Stable Diffusion could not function in the way that it presently does, had it not used Getty’s works (among many others) to train the model and generate its weights. AI models rely on “probabilistic matching” (the technical term for AI training) so inherently require access to a significant amount of copyright works.
On one hand, these copies must be reproduced for training (how else is ChatGPT to learn what a cat looks like?) but on the other, reproduction of a copyright work is one of the rights explicitly granted only to the copyright owner in the Copyright, Designs and Patents Act 1988 (CDPA). This brings us to Getty Images’ case against Stability AI.
Facts and Arguments
Getty Images, a renowned provider and licensor of photographs and other digital assets, brought an action against Stability AI, a UK-based AI company behind “Stable Diffusion”, a generative model for producing images based on user-input text-based prompts. It became evident to Getty that its images had been scraped, reproduced, and used to train the Stable Diffusion model after the model’s outputs were seen bearing Getty’s watermark. In some instances, the watermark was more distinct, in others blurred, or “distorted” to some degree.
An image generated using Stable Diffusion v1.2 (during experiments carried out by Prof. Hany Farid, an expert witness, and Emma Varty, a solicitor acting for Getty), using prompt “Obama, news photo”
Getty originally alleged trade mark infringement, passing off, database rights infringement, and copyright infringement (including claims for “primary infringement” under Section 16 of the CDPA, and “secondary infringement” under Sections 22 and 23).
Stability sought to strike out Getty’s copyright claims on the basis that Stable Diffusion was developed and trained entirely outside the UK, in U.S. Amazon Web Services data centres. In the UK, Stability merely made software available on its website and did not deal in any “tangible things”, and so (Stability argued) there could be no “infringing article” and hence no primary or secondary infringement. Getty countered that training must have occurred in the UK because its works were downloaded onto UK servers during the training of Stable Diffusion. But Stability maintained that no employees worked on the training and development of Stable Diffusion from the UK, something Getty conceded at trial, thereby abandoning one of its claims for primary copyright infringement. Getty abandoned its other claim for primary copyright infringement in response to Stability blocking users from entering certain input prompts to its service which had been identified as capable of generating infringing outputs.
Getty also abandoned its database rights infringement claim during proceedings, so in the end the key substantive issues to be decided at trial were trade mark infringement (which Getty largely won), passing off, and secondary copyright infringement.
Stability acknowledged that copyright-protected images bearing the Getty (and iStock) watermarks had appeared in the LAION-5B dataset used to train the Stable Diffusion model, but maintained innocence of secondary infringement on the grounds that:
a) The model could not be construed as an “article” within the meaning of Sections 22 and 23 CDPA because it was not a tangible thing, and
b) Even if the model were an “article” in this sense, it could not qualify as an “infringing copy” of any of the copyrighted works within the meaning of Sections 22 and 23 CDPA (as defined in Section 27 CDPA) because it is not a “copy” of these works at all – its internal weights and biases do not contain the original images on which the model was trained, and the original images are not present within, and cannot be reproduced from, the model.
Mrs Justice Smith found against Stability on point a) above, but found in their favour on point b), the focus of the commentary in this article.
“Infringing Copy”
Mrs Justice Smith’s refusal to find copyright infringement pivoted on the finding that the imported Stability models were not an “infringing copy”, and that while an infringing “article” may be intangible in its nature, Stable Diffusion itself was not an infringing copy. AI models, while trained on terabytes of data, only themselves consist of mere gigabytes, as their model weights merely reflect the statistical patterns learnt during training, and (according to Stability AI) never at any point contain any copyrighted work.
Therefore, taking the ordinary meaning of an infringing copy, Smith J’s approach and conclusion seem logical. However, the court may have erred in its interpretation of what is meant by an “infringing copy”.
The CDPA states at s.27(3) that:
“An article is also an infringing copy if it has been or is proposed to be imported into the United Kingdom, and its making in the United Kingdom would have constituted an infringement of the copyright in the work in question”.
This paragraph seemingly exists as a safeguard to prevent an unauthorised third party from carrying out an infringing act in a separate jurisdiction and then importing the resulting infringing copies into the UK. The manufacture of counterfeit DVDs, for example, would still be caught under secondary copyright infringement by s.27(3). In effect, the test creates a territorial extension of the UK’s IP protection, preventing parties from circumventing UK copyright law.
However, the facts of the case required the court to depart from the traditional interpretation of secondary infringement concerning physical goods crossing the border. Classic authorities on “importation” involve tangible items, e.g. shoes in LA Gear. [2] Even in Sony v Ball, [3] a RAM chip was held by Laddie J to be an infringing article albeit for only a short time, namely the split second the copy was held on the RAM. By contrast, Getty required the court to address intangible model weights distributed over the internet and accessed via hosted services. While Smith J distinguished past interpretations that focused on goods being physically received into the UK as central to the infringing act (as in LA Gear) and held that downloading an intangible article into the UK amounts to importation, she found that access and use of Stable Diffusion did not, as no copy ever entered the jurisdiction. In short, what it means to import an infringing article has stretched with modern technology to encompass more than physical goods, but not so far as to capture AI models. This is the public policy tension we see in Getty: how to protect copyright owners against extrajudicial infringement yet ensure that secondary infringement does not stretch to cover articles that would never constitute a copy of a protected work. In Getty, the court prioritised the latter. However, in doing so, it has exposed a gap for potential copyright infringement.
This is where the wording of s.27(3) becomes crucial, and the court’s interpretation likely erred. The word “also” in s.27(3) is significant. It suggests that what follows is an equal and alternative definition for an infringing copy. It follows that the process of “making” an article is equally as significant as its importation, whether or not the model ever contained a copy of the infringing works.
Smith J found that it was not enough that the “making of the copies of the Copyright Works coincides with the making of the Model”. Rather, due to the intangible nature of the model, that does not contain the works themselves, s.27(3) does not extend to include the digital creation of Stable Diffusion’s weights. This is arguably the primary error of the court and where hope for the Court of Appeal’s judgment lies.
Arguably, “making” cannot merely be the creation of the model weights. If one “makes” a cake, this act is not satisfied merely by taking the sponge out of the oven, but rather the selection, purchase and mixing of ingredients in specific ratios too. When training AI models the “ingredients” (i.e., copyright works) are used to create the “cake” (i.e., the model weights) but they do not exist themselves within the model (and cannot be reconstituted after baking). However, the training, were it to take place in the UK, would have constituted copyright infringement due to the reproduction of copyright protected works necessary to train the model, which Mrs Justice Smith acknowledges. The fine-tuning of model parameters is parasitic on this reproduction; the models must learn their patterns somewhere, and there’s a reason datacentres are being built at an exceptional rate.
This suggests that “making” in s.27(3) should not be limited to digital memorisation or storage in the model but must also include the extensive training that AI models have to undergo in order to generate output, training that relies on reproduction of copyright protected works. It follows that Stable Diffusion is therefore an infringing copy under s.27(3) as its making (in the U.S.) relied on the reproduction of works, which would constitute copyright infringement in the UK under the CDPA.
Memorisation and GEMA v OpenAI
The court’s finding that “an infringing copy must be a copy”, is based on the fact that the model weights “never contain or store those copies”. Getty did not dispute this at first instance so this point won’t be addressed by the Court of Appeal. However, the recent Munich Regional Court decision in GEMA v Open AI [4] may help to shed some light on this issue, looking forward.
In early November 2025, the Munich Regional Court, which specialises in copyright law, upheld the claims made by the German collecting society and performance rights organisation for music (GEMA) for injunctive relief and damages against OpenAI, the developers of famed J.A.R.V.I.S. impersonator, ChatGPT. The ruling focused heavily on the topic of ‘memorisation’, relying on scientific literature, finding that training data can become embedded in model weights and remain retrievable and therefore, reproducible. Following Article 2 of the EU InfoSoc Directive, which enshrines an author’s right to control reproductions of their work, the court understood the model weights to embody the training data, which the court termed ‘memorisation’ (‘Memorisierung’). Getty did not allege that Stable Diffusion stored copies of its copyright works, so cannot appeal on this point, however it is an interesting one and we will explore it further in a later article.
Reflecting on Getty’s case after the news of their permission to appeal is exciting, particularly as Getty intends to appeal the Court’s ruling on “making”, which we have explored in this article. For now, the fate of copyright owners’ rights and the legal status of AI training rests in the hands of the Court of Appeal.
References
[1] Getty Images (US) Inc v Stability AI Limited [2025] EWHC 2863 (Ch)
[2] LA Gear Inc v Hi-Tec Sports Plc [1992] FSR 121
[3] Sony Computer Entertainment Inc v Ball [2004] EWHC 1738 (Ch), [2005] FSR 9
[4] GEMA v OpenAI (42 O 14139/24, 11 November 2025)