Getty Images v Stability AI: implications for intellectual property, data protection and AI regulation

The UK High Court’s ruling in Getty Images v Stability AI [2025] EWHC 2863 (Ch), delivered on 4 November 2025, represents a pivotal moment in the legal treatment of generative AI models. The 205-page judgment grapples with the intersection of copyright law, trade mark protection, and the technical realities of machine learning technology. It provides useful guidance on how UK copyright law, which was originally conceptualised for a physical era, might apply to modern AI systems.

While the decision itself is UK-specific, it may be influential in shaping the global jurisprudential conversation around how AI models should be treated under law, particularly in terms of their inherent capacity, by reference to their natural operation, to infringe traditional intellectual property rights.

Getty Images' claims

Stability AI is the developer of an open-source model "Stable Diffusion", a latent diffusion model that transforms text prompts into images.

Getty Images, a provider of stock photography, made a range of claims against Stability AI, namely that:

• Stability AI had scraped approximately 12 million images and captions from Getty's site in order to train Stable Diffusion without a licence, thereby infringing copyright in those images (Training Claim);

• certain outputs generated by Stable Diffusion reproduced substantial parts of those scraped and copyrighted images, in infringement of Getty's copyright (Primary Claim);

• Stability AI’s distribution of Stable Diffusion facilitated downstream infringements by end users (End User Claim);

• Stability AI infringed Getty's copyright by importing into the UK, an "infringing article" (Stable Diffusion) (Secondary Claim); and

• in some cases, the content produced by Stable Diffusion also contained Getty’s watermark, forming the basis for an additional claim of trade mark infringement (Trade Mark Claim).

Getty sought a declaration from the Court that Stability AI's conduct infringed its copyright and trade marks, as well as an injunction to restrain Stability AI from further infringement.

Stability AI's response

Stability AI conceded that some Getty assets were present in the dataset used to train Stable Diffusion, but denied liability and sought reverse summary judgment to dismiss the Training Claim and the Secondary Claim, on the basis that:

• in respect of the Training Claim, the training of Stable Diffusion's underlying model did not occur in the UK, and therefore the Training Claim could not be made pursuant to the UK Copyright, Designs and Patents Act 1988 (CDPA); and

• in respect of the Secondary Claim, Stable Diffusion's underlying model is not an "article" that could infringe copyright under the CDPA, and accordingly, there was no basis for the Secondary Claim.

Justice Smith refused Stability AI's application for reverse summary judgment on the basis that there was, to the mind of the Court, a reasonable prospect that evidence could be uncovered through discovery, which would establish where Stable Diffusion was trained (with the Court also taking into account that Stability AI is UK-based).

However, Getty eventually abandoned the Training Claim, due to jurisdictional and evidential gaps – it could not adduce evidence that Stability AI's training of Stable Diffusion occurred in the UK. Similarly, it withdrew the End User Claim due to evidentiary issues.

As to the Primary Claim, the Court expressly preserved that claim for future determination, on the basis that the Primary Claim related to more garden-variety copyright infringement, which could be determined at a later stage in separate proceedings, if the parties so wished.

This left only the Secondary Claim and the Trade Mark Claim to be determined at trial.

The Secondary Claim

Under the Secondary Claim, the Court considered whether an AI model could be considered an "article" for the purposes of the CDPA, which would then inform the strength of Getty's infringement claim.

While Stability AI argued that the term "article" should be limited to tangible objects, the Court disagreed, adopting a purposive interpretation of the law based on technological neutrality. It held that software embodiments such as model checkpoints can indeed be "articles", reasoning that legislators could not have intended to exclude intangible media in an era of digital dissemination, and that prior case law supports the definition of "article" being interpreted to include such intangible media.

However, notwithstanding the Court's view, Getty was still required to demonstrate that Stable Diffusion (i.e. the underlying model) was itself an "infringing copy" of the works used to train it. Ultimately, the Secondary Claim failed, with the Court holding (assisted by expert evidence), that the model was not an "infringing copy" of the original works because:

• Stable Diffusion is a system of weights and parameters that, via a process of statistical calculation, merely draws semantic connections between data points within the body of training data, but does not, in fact, create perceptible copies of that training data by virtue of the function of the model; and

• as per Stability AI's argument, after training, the model no longer contains or stores the relevant training data.

This aspect of the Court's judgment could materially influence the prevailing perception of the nature of AI models and the extent to which they could be deemed to infringe copyright in their underlying training data corpora; because the Court ruled (at least in this instance) that the AI model, of itself, did not contain or store training data. Naturally, it may be that in the future, courts take a more liberal view on the extent to which an AI model could be considered to "contain" or "store" the works in its training data corpus, but for now this case has established a clear principle.

The Trade Mark Claim

Getty also claimed trade mark infringement, based on Getty’s registered trade mark (the "Getty Images" watermark) appearing in certain of Stable Diffusion's outputs.

Ultimately, the Court found in favour of Getty's argument that the mark was:

• reproduced in a context capable of denoting trade origin;

• used "in the course of trade"; and

• therefore infringing, because Stability AI had distributed the model commercially.

However, this claim only applied in relation to older models of Stable Diffusion, where such infringement was established to have occurred. Damages will be assessed in due course in respect of this successful aspect of Getty's claim.

Getty also alleged passing-off under common law, in connection with the same facts and circumstances giving rise to the alleged trade mark infringement. However, the Court rejected this argument on the basis that Getty had failed to adduce evidence of any sort of consumer confusion arising from the reproduction of Getty's trade marks in Stable Diffusion's outputs.

The effect on privacy-related thinking

This decision represents a clear view by the UK High Court that AI models – at least those with the configuration and parameters of Stable Diffusion – are not, of themselves, data repositories in the same way that more traditional databases are. This view stands in contrast with the European Data Protection Board’s (EDPB) 2024 opinion on AI models and personal data, which treated machine-learning models as potential "data stores" and warned of a potential re-identification risk, due to models having the capability to reconstruct personal data and re-identify individuals through model outputs.

Under this view, it may be necessary for data controllers to treat the operation of an underlying model as an exercise in personal data processing (and therefore require consideration of how the GDPR might apply to that operation).

Leaving aside the potential risk that may flow from the outputs produced by an AI model (for example, the re-identification of individuals through outputs containing personal information), this raises a fundamental question for courts and regulators as to whether the operation of an AI model can attract privacy risk. In other words, if, as per the UK High Court's decision, a model cannot be considered to "contain" copyrighted works, then how can it, at the same time (and as per the EDPB opinion), be said to "contain" personal data? Harmonising this view across copyright and privacy laws will be essential in ensuring a uniform jurisprudential approach in respect of the treatment of AI models.

For any stakeholder in the AI ecosystem however, this decision, as contrasted against that signalled by the EDPB, highlights the importance of understanding the nuances of the somewhat fragmented treatment of AI systems across different legal and regulatory regimes globally. It also illustrates how multiple, and potentially cumulative, obligations, definitions and enforcement mechanisms may apply across different regimes, such as privacy, intellectual property and consumer protection. In practical terms, forming a view on copyright in the AI context may not be definitive if privacy, trade mark, consumer protection or other regulatory obligations are overlooked.

Future implications

While the considerations involved in this decision were ultimately quite narrow (especially as the Primary Claim was reserved), it may prove impactful in influencing the attitudes and actions of a range of different stakeholders in the AI ecosystem, including regulators and lawmakers. Further, it has attracted attention beyond the UK in the context of the potential interaction between AI functionality and intellectual property infringement.

Rights holders

Evidentiary measures will continue to be important for copyright owners seeking to safeguard their works against AI model usage. In this regard, rights holders should consider implementing steps to easily establish and trace ownership of copyrighted works (for example, by the use of watermarks or hashing), which will assist with ongoing detection of unauthorised use. Further, the Training Claim (though ultimately relinquished by Getty) emphasises the importance of being able to establish the geographical attributes relating to any allegations of copyright infringement.

Developers

For developers, the Training Claim highlights the importance of the careful selection of the location of any technical acts performed in relation to the training of AI models, including caching and pre-processing. Developers should carefully evaluate and determine the geographical reach of their AI-related activities (including their cloud-location strategies).

Of course, the decision could embolden AI developers and incentivise avoidance-type behaviours (such as training AI models in more "accommodating" jurisdictions, to avoid the application of certain laws). Developers may also seek jurisdictions that adopt more-favourable approaches to the use of data scraped from the internet, including in relation to licensing or compensation due to rights holders.

It is possible that different jurisdictions will adopt varied or nuanced policy positions with respect to AI development (including training models), in seeking to balance the desire to promote investment and innovation against appropriate legislative and regulatory protection for rights holders. Recently, the Australian Attorney-General signalled that the Federal Government would not pursue a text and data mining exception to Australian copyright law (which exception would allow for the training of AI models on scraped data without infringement), but that it would conduct a public consultation on different proposed licensing models for AI's use of copyright content in due course.

However, we consider that, given the industry backlash we have seen around the world in recent years, as well as increasing calls for legislative reform, it is more likely that there will be a move toward greater legislative and regulatory protection for rights holders, and more fulsome obligations on developers designed to avoid infringement (for example, mandated licensing).

Irrespective of the form of any future regulation, it will be important for the use of AI models to be supported by the adoption of strong governance practices, the implementation of appropriate risk controls and reliable record-keeping with respect to how data is treated.

Commercial considerations

From a commercial perspective, this decision may also prompt certain rights holders to develop voluntary licensing programs for their content, as a pre-emptive move to capture the value in that content. We are already seeing this approach in a number of foreign markets, with rights holders pro-actively developing and advertising remuneration models for the licensing of their content.