Getty Images v Stability AI: the UK Case, Status and Rulings

Getty Images sued Stability AI in the High Court of Justice in England and Wales in January 2023, in what was expected to be the first full-trial test of generative AI training against UK copyright law. By the time judgment was handed down on 4 November 2025, it had become a case about something else entirely, and the reason is more instructive than the result.

What was originally claimed

The claim, brought in the Business and Property Courts, ran on several independent tracks: that training Stable Diffusion on Getty’s images infringed copyright; that outputs reproduced substantial parts of Getty’s works; that importing and dealing in the trained model in the UK amounted to secondary infringement under sections 22 and 23 of the Copyright, Designs and Patents Act 1988; infringement of database right; and trade mark infringement and passing off, based on Getty and iStock watermarks appearing in generated images.

Stability applied for reverse summary judgment and strike-out on parts of the claim in 2023; the court allowed the case through to trial, and the trial took place in June 2025 before Mrs Justice Joanna Smith. Judgments of the English courts are published at the National Archives Find Case Law service, which is where the text should be read rather than any summary, including this one.

The claims that fell away at trial

During and after the trial, Getty abandoned the training and output copyright claims. The reason was territorial and evidential rather than doctrinal: UK copyright law reaches acts of copying done in the UK, and Getty was unable to establish on the evidence that the relevant training copies had been made here. The development and training happened elsewhere, on infrastructure elsewhere, and the disclosure and witness evidence did not put those acts within the jurisdiction.

This is the most transferable lesson in the case and it has nothing to do with AI. A claimant in a national court must prove infringing acts within that court’s reach, and for model training those acts are opaque, distributed and frequently offshore. It is the same evidential wall that shapes the practical value of the EU’s rights reservation mechanism, where the reservation is easy to state and the infringement hard to demonstrate.

What the judgment decided

What was left for judgment was the secondary infringement claim and the trade mark claim.

• Secondary infringement failed. Sections 22 and 23 of the 1988 Act concern dealing in an “article” which is an infringing copy. The court held that the model’s weights are not an infringing copy of the training images, because the model does not store or reproduce the works — it stores learned parameters. On that reasoning an article cannot be an infringing copy of works it does not contain, and importing or making the model available in the UK therefore did not engage those sections.
• Trade mark infringement succeeded, narrowly. The court found infringement in relation to watermarks appearing in generated outputs, but the finding was limited in scope and largely historic, attaching to particular model versions and particular circumstances rather than to the product generally.

The weights holding is the part with reach beyond the parties. It answers a question that had been genuinely open in English law — whether a trained model is itself an infringing copy of what it was trained on — in the negative, on the evidence about how these models store information. That reasoning matters for anyone importing or distributing model files in the UK, and it points in a different direction from the more nuanced position the US Copyright Office took on memorisation, set out in the Part 3 report page.

What the judgment does not decide

This paragraph is the one most worth taking away, because the case is widely cited for propositions it does not support.

• It does not decide that training on copyrighted works is lawful in the UK. That claim was withdrawn. The court made no finding on it, and section 29A remains the only relevant exception.
• It does not decide that outputs cannot infringe. The output copyright claim was also withdrawn. A model that reproduces a substantial part of a work in an output remains exposed on ordinary principles.
• It does not decide that models never contain copies.The holding rests on evidence about how these particular models store information. A system that did retain retrievable copies would present a different question.
• It does not resolve UK policy. The consultation described in the UK AI and copyright consultation page is the process that will change the law, if anything does. A judgment on the law as it stands does not pre-empt it.

Whether the judgment is appealed, and what an appellate court makes of the secondary infringement reasoning, will determine how durable the weights holding is. Check Find Case Law for subsequent history before relying on it.

The parallel US proceedings

Getty also brought proceedings against Stability AI in the United States, in the District of Delaware. That is a separate action under different law, with a fair-use defence that has no counterpart in English law, and nothing decided in London disposes of it. Do not read reporting of the two interchangeably; the English case turned on territoriality and the meaning of “article” in a 1988 statute, neither of which is a question an American court will ever ask.

More broadly, the divergence between jurisdictions on the same underlying technology is now the defining feature of this area. English law asks where the copying happened and whether the thing distributed is a copy; American law asks whether the use was fair, as in the New York Times proceedings; and EU law asks whether an exception applied and whether rights were reserved. A model developer operating across all three is exposed to three unrelated analyses of the same act, and a favourable outcome in one says very little about the others.