Getty Images v Stability AI: What the UK's First AI Copyright Ruling Actually Decided
The UK High Court's landmark ruling in Getty Images v Stability AI didn't settle whether training AI on copyrighted works infringes. Here's what it actually held, and what firms should tell clients.
On 4 November 2025, the High Court handed down judgment in Getty Images (US), Inc. and others v Stability AI Ltd, [2025] EWHC 2863 (Ch). It is the first substantive UK court decision to grapple with generative AI and copyright, and it is still shaping how firms advise clients on both sides of the AI supply chain. It is also widely misdescribed. If you’ve seen it summarised as “UK court says AI training doesn’t infringe copyright,” that headline is doing more work than the judgment actually did.
What the Case Was Actually About by the Time It Reached Trial
Getty’s original claim was broad: that Stability AI had unlawfully used millions of Getty owned and licensed images to train its Stable Diffusion model, infringing copyright, database rights and trade marks. By the time the case reached trial, Getty had withdrawn its primary copyright and database infringement claims entirely. That’s a significant fact that gets lost in most summaries. The court was not asked to decide, and did not decide, whether training a model on copyrighted images without a licence is itself infringement under UK law. That question remains open.
What was left for the court to decide was narrower: whether distributing the model’s weights amounted to secondary copyright infringement, and whether outputs that displayed visible Getty or iStock watermarks infringed Getty’s trade marks.
What the Court Held
On the secondary infringement claim, the court found that model weights are not “infringing articles” for the purposes of UK copyright law. Weights represent statistical relationships learned during training, not a stored copy of any particular image, and the judge held that an infringing article must “at some point embody, contain or store” the protected work. Because the weights don’t do that, distributing them doesn’t amount to distributing infringing copies.
On the trade mark claim, the court reviewed hundreds of example outputs and found only three instances of actual infringement, all involving outputs that reproduced a Getty or iStock watermark. That is a narrow, fact specific result, not a finding that Stable Diffusion systematically infringes Getty’s marks.
What Firms Should Actually Tell Clients
The temptation is to treat this as a green light for AI training generally. Resist it, for three reasons.
First, the underlying question of whether training itself infringes was never decided because Getty dropped that claim, so this judgment doesn’t answer it, in the UK or anywhere else. Second, the ruling turns on a specific technical finding about how diffusion models store information, and that finding won’t automatically transfer to every AI architecture your client might be evaluating or building. Third, the UK still has no text and data mining exception for commercial AI training of the kind being debated in other jurisdictions, which means the legal position for a client actually building or fine-tuning a model in the UK remains genuinely unsettled, whatever the headlines suggest.
For clients licensing or deploying third party AI tools, the practical read across is narrower still: a vendor cannot point to this case as a general indemnity against copyright claims, and firms should keep asking vendors the same due diligence questions about training data provenance they were asking before the judgment. For clients whose content might have been used to train a model, this case is a reminder that a secondary infringement claim built around model weights is now a harder route in the UK, but it does not close off other routes, and it says nothing about claims outside the UK’s jurisdiction.
Why This Matters Beyond the UK
Firms advising in India and Singapore should treat this less as binding precedent and more as an early data point in a live global debate that will keep shifting as parallel cases move through US and EU courts. The UK’s approach, deciding narrowly on the specific technical mechanism rather than the broader training question, is itself instructive: expect other courts asked the same question to reach for equally narrow, fact bound answers rather than sweeping rules, at least for now.
The tool is not the advantage here either. The firm that reads past the headline to what was actually litigated, and what was quietly dropped before trial, is the one giving its client an answer that will still be accurate in six months.