.svg?rev=a492cc1069df46bdab38f8cb66573f1c&hash=2617C9FE8A7B0BD1C43269B5D5ED9AE2)
At a Glance
- The English High Court has delivered its keenly anticipated judgment in Getty Images v Stability AI.
- While several important issues were abandoned in the run up to and during the trial, the court reached a position on a number of points with wider implications for the question of how to balance the competing interests of AI developers and the creative industries.
- Getty’s claim of secondary copyright infringement failed on the basis that Stable Diffusion’s model weights did not constitute infringing copies. However, the court held that there had been some limited historic trade mark infringement in the outputs generated by early versions of Stable Diffusion.
Background to the Case
Stability AI (Stability) carries on business in the field of machine learning software, including deep learning models for image and music generation, and large language models for the generation of text output. Getty Images (Getty) is a visual media company which has a substantial library of stock images, video and other forms of media assets, including the iStock website.
The UK litigation is part of a broader set of claims brought by Getty against Stability in multiple jurisdictions. The claims centred on Stability’s AI model, known as Stable Diffusion, which is a latent diffusion model, meaning that it makes use of model parameters, known as model weights, to transform prompt text or images into output synthesised images through sampling from a probability distribution modelled on training data. These model weights are iteratively updated through the training process, which involves repeated exposure to significant quantities of data, generally scraped from the internet.
Stable Diffusion was at least partially trained on Getty’s stock of visual assets (among other websites) without consent. Getty sought a declaration that Stability had infringed Getty’s copyright, trade mark and database rights, and an injunction to restrain Stability from any further infringement. The claims were first brought in January 2023, and the first trial on the issue of liability took place in June 2025.
The judgment was eagerly anticipated, and as the judge noted: “[b]oth sides emphasise the significance of this case to the different industries they represent: the creative industry on one side and the AI industry and innovators on the other. Where the balance should be struck between the interests of these opposing factions is of very real societal importance”. However, Getty’s decision to drop the primary copyright infringement claims was a blow to AI developers and copyright holders alike who were expecting that the outcome of this case would provide greater clarity on the interaction between UK copyright law and AI. As the judge noted, the court could only determine the issues arising from the (diminished) case that remained before it.
Primary Infringement: Training and Development Claim
Getty claimed that the training and development of Stable Diffusion amounted to primary infringement of copyright (i.e., through copying and communicating copies to the public (under the Copyright Designs and Patents Act 1988 (CDPA) Section 17 and 20)). However, there was no evidence that training and development of Stable Diffusion took place in the UK. Therefore, the Training and Development Claim, which was a key line of attack for Getty, was abandoned.
Primary Infringement: Outputs Claim
Getty also brought claims relating to 17 copyright works (out of the millions allegedly used to train Stable Diffusion). Stable Diffusion took action during the course of the proceedings to block certain prompts which had been alleged to generate infringing output, which the court ruled enabled Getty to largely achieve any relief to which it would have been entitled. As a result, the outputs claim (together with inherently linked claims for database rights infringement) was also abandoned.
Secondary Copyright Infringement — Background
Getty also advanced claims of secondary copyright infringement. Under the CDPA, acts of secondary infringement of copyright include:
- importing an infringing copy — i.e., importing an article which is an infringing copy (section 22 CDPA); or
- possessing or dealing with an article which is an infringing copy (section 23 CDPA).
The CDPA does not define an “article”, but the Court held (after a detailed analysis of the principles of statutory interpretation) that it can include electronic copies stored in an intangible format (e.g., software stored on Amazon Web Services).
Getty did not seek to argue that Stable Diffusion was itself a copy of, or that it stored within it any copies of, the copyright works, rather that, pursuant to section 27(3) CDPA, Stable Diffusion was an infringing copy because the making of its model weights would have constituted infringement of the copyright works had it been carried out in the UK.
The court held on the facts that the model weights did not store or reproduce the original works themselves, but rather were a product of the training process which made use of the patterns and features of the stock of images. Therefore, the articles that were being imported (Stable Diffusion’s model weights) were not considered infringing copies, notwithstanding that this leaves open the possibility that there could have been infringement during the making of the model weights had the relevant development and training occurred in the UK.
Trade Mark Infringement
Getty succeeded in claims that were both “historical and limited in scope” in respect of trade mark infringement relating to specific outputs from early versions of Stable Diffusion that generated synthetic images with “Getty” or “iStock” watermarks. Getty claimed that normal use of Stable Diffusion by users in the UK would in some cases, generate synthetic images bearing Getty’s trade marks, contrary to sections 10(1), 10(2) and 10(3) of the Trade Marks Act 1994 [10].
To succeed, Getty would have needed to establish that at least one output which included the Getty / iStock watermarks had been generated in the UK. While Stability had accepted in principle that Stable Diffusion could be used to generate watermarked images, Getty failed to show that such images had been generated “in the wild” as opposed to in only a handful of hand-picked contrived experimental prompts carried out for the purposes of the litigation.
Broader claims under section 10(3) (unfair advantage, dilution and tarnishment) failed due to lack of evidence of economic harm or reputational impact. The court found no proof that users sought out watermarked images nor that the presence of watermarks changed the behaviour of the average consumer. Furthermore, the court declined to address a claim of passing off, stating it added nothing beyond the trade mark findings.
Key Takeaways
The judgment was eagerly anticipated in that it is the first major English law case which provides a detailed analysis of how UK copyright law applies to AI. Getty’s evidential difficulties in establishing that any of the allegedly infringing acts took place in the UK and withdrawal of many of their claims mean that a number of issues are left unanswered.
Creative industries will take some comfort from emphasis in the judgment that AI developers have “absolute control” over training data and bear responsibility for outputs, rejecting the notion that they are merely neutral tool providers. While the judgment holds developers responsible for dataset choices and output filtering, there remains debate on the extent of liability for downstream users.
However, the key findings relating to secondary infringement — that model weights are not infringing copies in themselves since they do not store or reproduce the original copyright works used as training data — are likely to shape the course of subsequent litigation in this area. To establish infringement under UK law, claimants will need to show that there has been actual copying of training data in the UK. This will be highly fact-specific and will require a detailed analysis of the relevant model and its training data in each case. It remains to be seen how parallel proceedings in the United States shape the international landscape for AI developers.
The Broader Policy Debate
The UK’s recently introduced Data (Use and Access) Act 2025, has (for the time being) sidestepped the broader policy issues in respect of the use of copyright works in AI training datasets. Instead, as set out in our earlier update, it mandates the Secretary of State for Culture, Media and Sport to publish an economic impact assessment of the following four policy options previously identified by the UK government, taking into account the impact of those options on copyright owners and developers of AI systems:
- Express licensing requirements for training AI
- A broad data-mining exception
- A data-mining exception allowing rights owners to reserve their rights in respect of access by AI — i.e., an opt-out right (which was the UK government’s preferred option, but has been heavily criticised and lobbied against by creative industries)
- No change to the current system
The Act also requires a detailed report to be used on the use of copyright works in AI development, taking into account licensing arrangements, enforcement mechanisms and technical standards.
