Getty v Stability AI: The Core Question on AI Training Remains Unanswered

The High Court has ruled in Getty Images v Stability AI, but businesses should not mistake a judgment for clarity. While AI developers secured an important technical victory, the decision leaves the most commercially critical question unresolved: is training an AI model on copyrighted material lawful in the UK?

Background

In 2023, Getty Images launched legal proceedings against Stability AI, the developer of the text-to-image generative AI model Stable Diffusion. Getty alleged that Stability had infringed its copyright and trade mark rights by using millions of Getty’s images to train the model without authorisation. It further argued that the trained model itself amounted to an “infringing copy” under UK law, and that some of Stable Diffusion’s outputs even reproduced Getty’s distinctive watermarks.

The case drew significant attention as the first major UK decision to consider how copyright and trade mark laws apply to generative AI models. On 4 November 2025, Mrs Justice Joanna Smith DBE largely found in favour of Stability AI, dismissing Getty’s copyright claims but upholding a limited finding of trade mark infringement.

The Decision: What the Court Did (and Didn’t) Decide

The Court rejected Getty’s core allegation that the Stable Diffusion model amounted to secondary copyright infringement under sections 22, 23 and 27 of the Copyright, Designs and Patents Act 1988 (CDPA). Getty had argued that by making Stable Diffusion available for download in the UK, Stability was importing or dealing with an “article” that was an “infringing copy” of Getty’s works.

The Court disagreed. Justice Smith found that AI model weights, the statistical parameters determining how the model processes data, do not store or reproduce the visual information contained in training images. Instead, the model captures learned patterns rather than stored copies. As a result, Stable Diffusion could not be considered an “infringing copy” under the CDPA, and its distribution in the UK did not amount to secondary copyright infringement.

This is an important clarification for AI developers: where a model’s architecture does not store or reproduce copyright works, the model itself may not be treated as an infringing article under UK law.

The Court also addressed what constitutes an “article” for the purposes of secondary copyright infringement. Historically, “article” was understood to refer to tangible, physical objects. However, Justice Smith held that “article” is broad enough to include intangible information such as software or AI model weights, particularly where such copies can be stored electronically in cloud computing. The Court reasoned that Parliament intended the statute to adapt to modern technology, and that limiting “article” to tangible objects would deprive copyright owners of protection where works are dealt with electronically. This broader interpretation may diverge from EU law, which traditionally limits secondary infringement provisions to tangible objects.

But here is the critical gap. The Court did not (and could not) decide whether the act of training an AI model on copyrighted material in the first place constitutes primary copyright infringement. Getty had initially brought a claim for primary infringement, alleging that copying its images during the training process infringed its rights. That claim was withdrawn because Getty could not produce sufficient evidence that the training took place within the United Kingdom. Stability developed Stable Diffusion abroad, and UK courts have no jurisdiction to rule on acts of copying carried out outside the UK.

This means the central legal question for the AI industry remains unresolved. Developers do not know whether training on copyrighted works without permission is lawful. Content owners do not know whether they have a viable claim. End-users may not know whether the tools they rely on are built on legally sound foundations. The judgment provides no precedent on this issue, and the uncertainty creates ongoing commercial and legal risk for everyone in the AI value chain.

Limited Trade Mark Infringement: Outputs Still Liable

Getty argued that Stability AI had infringed its registered trade marks because outputs from Stable Diffusion sometimes reproduced features identical or similar to its watermarks, causing consumer confusion about origin or endorsement. Stability countered that any appearance of watermarks was incidental, a by-product of the training data, and that end-users, not Stability, were responsible for the outputs.

The Court rejected this defence. It was held that Stability, as the party with meaningful control over output generation, was responsible for trade mark use under sections 10(1) and 10(2) of the Trade Marks Act 1994. The Court found limited infringement in relation to earlier model versions where watermarks appeared with sufficient clarity to cause consumer confusion. The infringement was limited because later versions reduced watermark reproduction.

Even if your model is structurally sound and not an “infringing copy,” you remain liable for what it produces. AI developers cannot claim immunity as passive distributors. Liability arises at the output stage, and a key mitigation defence is demonstrable, documented filtering and safety measures to prevent reproduction of protected marks.

What This Means for Your Business

The Court’s silence on primary infringement creates a legal vacuum that businesses must navigate strategically and proactively.

For AI Developers and Deployers

You should understand and document your model’s architecture. The Court’s finding that Stable Diffusion was not an infringing copy relied on detailed expert evidence showing that the model does not store or reproduce the images used in training. Commission technical and legal assessments now to prove your model functions the same way. This evidence will be critical if you face secondary infringement allegations in future.

You should also know where your training occurs. Territoriality determined this case. If you train models outside the UK, you may reduce exposure to UK primary infringement claims, but this does not insulate you from claims in the US, EU, or other jurisdictions where the legal position may be different or more hostile. Map your jurisdictional exposure and design your development processes accordingly.

Finally, you should implement robust output filtering. The trade mark finding confirms that outputs are a liability even when the model itself is not infringing. Build, maintain and document filtering mechanisms to prevent reproduction of watermarks, logos and other protected marks. Make this a core part of your product design, not an afterthought, and ensure you can evidence your efforts in any dispute.

For Content Owners and Rights-Holders

Your most reliable protection is now contractual. Review and update your terms and conditions, licensing agreements and contributor contracts immediately. If you license content, you may want to include clear provisions requiring consent or separate payment for AI training use. Do not assume existing clauses cover AI.

You should also monitor for trade mark infringement in AI outputs. The Court confirmed that watermark replication can constitute actionable infringement. Implement monitoring systems to detect unauthorised reproduction of your brand identifiers in AI-generated content and enforce aggressively when you find it. Watermarking your content is now a defensive necessity, not an optional branding choice.

Finally, consider jurisdiction strategically. The UK route to primary infringement claims may be limited for now.

For Businesses Using AI Tools

If your business uses generative AI for marketing, design or internal workflows, you should vet your vendor contracts. You do not want to be solely liable if an AI-generated image in your advertising campaign is later found to infringe a third party’s copyright or trade mark.

You should also establish internal compliance policies. Implement clear, enforceable rules governing how employees use generative AI tools. Require checks on outputs for potentially infringing content. Document your compliance measures. This can reduce both your legal liability and your reputational risk if something goes wrong.

Conclusions and Next Steps

This judgment does not give a green light for training AI models on copyrighted works. It confirms that proving an AI model is an “infringing copy” is technically and legally difficult, but it does not resolve whether training itself is lawful. The law is struggling to keep pace with technology, and businesses cannot afford to wait for clarity that may take years to emerge.

Whether you are developing AI models, licensing content, or using AI tools in your operations, the most effective protection lies in proactive strategy. Our intellectual property and technology lawyers work with businesses at every stage of the chain to audit contracts, assess risk exposure, and build IP strategies tailored to your business model and risk appetite. Contact us to discuss how we can support you.