Getty v Stability AI: technical knockout goes to appeal

Permission has now been granted for Getty Images to appeal part of the High Court’s decision in Getty Images v Stability AI. The appeal is narrow, technical and, on its face, dry. Its implications are anything but. At stake is whether UK copyright law can treat a trained AI model as an ‘infringing copy’ for the purposes of secondary infringement, even where the model does not store recognisable copies of the works on which it was trained.

That question goes well beyond this dispute. It raises a broader issue about how copyright law should apply to AI systems that don’t look like traditional copies, but can still reproduce protected works.

What is (and is not) being appealed

We described the High Court’s November 2025 judgment as a ‘technical knockout’ on copyright. Getty’s primary infringement and database claims had already fallen away by the end of trial, leaving a secondary infringement case based on the distribution and use of the Stable Diffusion model in the UK.

The court rejected that claim. Central to its reasoning was the idea that secondary infringement under the Copyright, Designs and Patents Act 1988 is concerned with dealings in infringing copies, and that an AI model which does not itself contain copies of copyright works cannot satisfy that requirement.

Getty has now been given permission to appeal on that point of statutory construction. Importantly, this is not an appeal about whether Stability AI actually copied Getty images during training, nor a belated attempt to argue that Stable Diffusion contains Getty’s images in any literal sense.  It is a legal argument about what Parliament meant by ‘infringing copy’, and how that concept should apply to intangible digital outputs like trained models.

The statutory fault line

The appeal ultimately turns on the meaning of a defined statutory term, ‘infringing copy’.  One provision in particular, section 27(3) of the CDPA, brings the issue into sharp focus.  In simplified terms, it provides that an article may be treated as an infringing copy if it is imported into the UK and its making in the UK would have infringed copyright.

That wording is doing something deliberate.  It shows that Parliament did not intend ‘infringing copy’ to be confined to objects that straightforwardly embody protected works.  Instead, the statute allows liability to turn, at least in some circumstances, on how an article was made, rather than solely on what it visibly contains.  Section 27(3) therefore operates as a statutory illustration of a broader point about the scope of the concept.

The difficulty arises when that logic is applied to modern technologies.  A trained AI model is not a physical object, nor a digital file that reproduces a work in any conventional sense.  Its weights encode statistical relationships learned from data, rather than storing images or text as such.  The question is whether the statutory concept of an ‘infringing copy’ is sufficiently flexible to accommodate that kind of object.

The High Court took the view that it is not.  Even taking section 27(3) into account, the court treated an ‘infringing copy’ as something that must still be a copy in the ordinary sense, i.e. an article that reproduces the protected work itself.  The thrust of Getty’s appeal is that this approach reads the statutory language too narrowly.

If the Court of Appeal accepts that argument, the consequences could be significant.  It would open the door to treating a trained AI model as an infringing article because of the process by which it was created, rather than because it contains recognisable copies of individual works.  That would mark a shift in emphasis from internal structure to provenance.

Related questions are emerging elsewhere. In GEMA v OpenAI, the Munich Regional Court found that reproducing copyrighted song lyrics in response to user prompts could amount to copyright infringement, focusing on the output of the system rather than the legal status of the model itself.  Although the legal framework differs from UK law, the decision underlines a shared difficulty for courts: applying copyright concepts built around acts of reproduction to generative systems whose legal effects arise at the point of use.

Why this matters in practice

The territorial nature of copyright law creates a practical problem for rightsholders challenging AI training.  Training often takes place outside the UK, using globally sourced datasets, long before a model is made available to UK users.  Establishing primary infringement tied to UK acts of copying can therefore be difficult.

Secondary infringement provisions offer a different route.  They focus on what happens when an article is imported, distributed or otherwise dealt with in the UK.  If a trained model can be characterised as an infringing copy because of the way it was made, then its deployment or distribution in the UK becomes legally significant, even where the training itself occurred elsewhere.

That is why this appeal matters, despite its narrow procedural scope.  It does not reopen the factual record in Getty v Stability. Instead, it asks a more fundamental question: whether UK copyright law is prepared to treat trained AI models as legally consequential digital products in their own right, rather than as containers whose legal status depends entirely on whether they visibly contain copies of protected works.

The uncomfortable question of ‘copies’ in AI models

AI companies frequently explain training in reassuring terms: models learn patterns, they do not store copies.

At a high level, that description is accurate.  A trained model does not contain image files or text documents that can be accessed, browsed or searched in the way conventional copies in a database or archive can.  But that framing has become increasingly difficult to treat as the end of the story.

A growing body of research shows that generative models can, in certain circumstances, reproduce training material far more closely than the ‘pattern learning’ reference suggests.  Studies of large language models document instances of verbatim memorisation, where models output long sequences of text that closely match material from their training data when prompted in particular ways.  Work on image generation has reached similar conclusions, demonstrating that diffusion models can sometimes emit recognisable training images, including individual copyrighted examples, under controlled conditions.

None of this means that AI models operate like hidden repositories of copyrighted works, or that reproduction is inevitable or routine.  But it does complicate the neat distinction often drawn between ‘learning from’ and ‘copying’ protected material.  From a legal perspective, the difficulty is that copyright has never been limited to obvious, file-based copies.  It has long been concerned with indirect, transient and technical forms of reproduction.

That tension sits uncomfortably with a legal framework that asks whether something is a copy by reference to how it looks internally, rather than what it can do in practice.  A model may not contain any identifiable reproduction of a work, yet still be capable (at least in some cases) of generating outputs that closely track protected material.  Whether that capability should matter for copyright law is far from settled.

This is not a question the High Court was required to resolve in Getty v Stability, and it is unlikely to be decided directly on appeal.  But it forms part of the wider background against which arguments about ‘infringing copies’ are now being made.  As research continues to erode the simplicity of the ‘no copies in the model’ narrative, courts may become less willing to treat internal structure alone as the decisive factor.

A case arriving at the wrong moment (or the right one)

The appeal comes at a time when the UK government is under pressure to recalibrate its approach to AI and copyright.  Earlier proposals that appeared to favour broad text and data mining freedoms have met resistance from rightsholders, and ministers have signalled a desire for a more balanced solution involving licensing, transparency and technical safeguards.

At the same time, the EU is moving ahead with a regime that places explicit copyright-compliance and transparency expectations on providers of general-purpose AI models.  US courts, meanwhile, are producing a patchwork of decisions that resist simple narratives about fair use, while increasingly forcing parties to grapple with the factual realities of training data and model behaviour.

Against that backdrop, the Court of Appeal decision could have influence beyond these immediate facts.  It may shape how aggressively UK rightsholders pursue secondary infringement theories, how AI developers assess distribution risk, and how policymakers think about the relationship between copyright doctrine and modern machine learning.

What to watch next

The appeal may not produce a sweeping declaration about the legality of AI training.  But it may do something subtler and more important: clarify whether UK copyright law is prepared to treat trained AI models as legally significant outputs of a copying process, even when they do not look like copies.

If the Court of Appeal endorses a broader reading of section 27(3), the centre of gravity in UK AI copyright disputes may shift.  Questions about training data, licensing and model deployment will become harder to compartmentalise.  And the comforting refrain that ‘the model doesn’t store copies’ may carry less weight.

For now, Getty v Stability AI remains a case about statutory interpretation.  But it is also a sign of where the real pressure points in AI copyright litigation are emerging and why they are unlikely to go away.

Humphreys Law are specialists in copyright and trade mark licensing and enforcement.  If you have any questions or issues concerning the above, contact a member of our team.  This piece was written by Tristan Morse.

All the thoughts and commentary that HLaw publishes on this website, including those set out above, are subject to the terms and conditions of use of this website.  None of the above constitutes legal advice and is not to be relied upon.  Much of the above will no doubt fall out of date and conflict with future law and practice one day.  None of the above should be relied upon.  Always seek your own independent professional advice.

Humphreys Law