On 4 November 2025 London’s High Court decided the first UK case on generative AI and intellectual property. Getty Images dropped its core copyright claims for lack of evidence of UK-based activity, won on three images, and was ordered to pay 69.4% of Stability AI’s costs. The lesson is economic: proving where a model was trained costs more than winning.
On 16 and 17 December 2025, in a procedural hearing almost nobody covered, London’s High Court did the arithmetic on a case that had run for nearly three years. Getty Images — the stock photography group that sued Stability AI in January 2023 over the unauthorised reproduction of roughly 12.0 million images, captions and metadata from its websites — was ordered to pay 69.4% of Stability’s legal costs, with interest, plus an immediate payment on account of £4,395,656.89. It had won on three images.
The artificial intelligence at the centre of the dispute is Stable Diffusion, a text-to-image model: a neural network trained on vast collections of image-caption pairs scraped from the web, which after training stores not the photographs but a set of numerical parameters — the weights — capable of generating new images from a prompt. Stability AI never denied that Getty images appeared in its training datasets. The legal question, and more importantly the economic one, was different: are those weights, distributed and hosted in the United Kingdom, an infringing copy of the photographs that shaped them?
What the judge decided, and what she did not
Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch), handed down by Mrs Justice Joanna Smith on 4 November 2025, is the first British ruling on copyright, trade marks and generative AI. It is also far narrower than its headlines.
Getty dropped its primary copyright and database right claims — the heart of the action, the ones about scraping and training — shortly before closing submissions, for want of evidence that those activities had taken place in the UK. The court therefore proceeded on the uncontested assumption that training had happened elsewhere. It did not rule on the territorial question: there was nothing left to rule on. Anyone who read the judgment as a holding about the geography of training liability read something the text does not contain.
What survived was the secondary infringement claim, which failed on reasoning that will matter: the model’s weights do not and never did contain reproductions of the works, and accessing a remotely hosted model — unlike downloading it — is not “importation” into the UK. But the principle should not be stretched beyond the proven facts. The same court held that an “article” under the Copyright, Designs and Patents Act 1988 can be intangible, so software or a model could in principle fall within the category. That point is contested — some legal analyses note it sits awkwardly with the scope of the EU-derived distribution right, and the court did not address the tension — and Getty obtained permission to appeal on precisely this issue, as a novel and important pure question of law. The appeal hearing is expected in November 2026. Nothing here is settled.
A win that does not cover the bill
Getty did win on trade marks, barely. Infringement was found only under sections 10(1) and 10(2) of the Trade Marks Act 1994, on a handful of outputs from historic model versions: the ISTOCK mark on an image generated by a 1.x release, the GETTY IMAGES mark on one from version 2.1. Concretely: two images infringing under both provisions, a third under 10(2) only. No infringement was found for SD XL or version 1.6, because no UK generations were proven. The section 10(3) claim — dilution and detriment to the marks’ distinctive character, potentially the most expensive head of loss for Stability — was dismissed because the scale of the phenomenon could not be measured. Additional damages were refused; passing off was never examined. Quantum on the marks has been referred to an inquiry that remains open.
Add it up. Three images, models already withdrawn from the market, no additional damages, no ruling on the core of the dispute, a voluntary undertaking from Stability instead of the penal notice sought, and 69.4% of the other side’s costs. Getty’s chief executive Craig Peters told CNBC in May 2025 that the company was spending “millions and millions of dollars” on the litigation. No public total exists, and none is needed: the ratio between spend and outcome is visible without one.
Why the economics of proof beat the doctrine
The economic lesson of the case is not about copyright. It is about what the burden of proof means when applied to a distributed system. To show that a machine learning model was trained in a given jurisdiction, you have to document where the compute cluster ran, where the data was downloaded, who executed which jobs and when. All of that sits inside the defendant company, spread across cloud infrastructure that by design has no single location. An image agency can count every photograph in its catalogue to accounting precision and still be unable to say in which country it was read.
Hence the first market implication: litigation risk around generative models is lower than rights holders hoped — not because the law protects them less, but because evidence is expensive. And the cost is asymmetric. Defending a model requires only disputing where training happened; attacking it requires reconstructing that fact. Every future case starts from this imbalance.
The second implication concerns Europe, and needs stating precisely: the English ruling is a first-instance national decision, in a system that has left the European Union, and binds no EU court. In the EU the territorial lever is legislative rather than judicial. Article 53(1)(c) of the AI Act requires providers of general-purpose AI models to put in place copyright compliance policies, and recital 106 asserts that those obligations extend to text and data mining carried out outside the Union. Whether that extraterritorial reach holds is debated in scholarship, precisely because it appears in a recital rather than in the article itself. But the design is the inverse of what emerged in London: instead of asking the rights holder to prove where training occurred, it asks the provider to answer for its own upstream supply chain, wherever it sits.
What Getty was doing while it was losing
The most interesting part is not in court. Getty’s second-quarter 2026 revenue came in at $229.1 million, down 2.5% year on year, with the creative segment at -2.6%. Its merger of equals with Shutterstock, announced in early January 2025, was referred to a Phase 2 investigation by the UK Competition and Markets Authority on 3 November 2025; the CMA’s final report of 15 May 2026 required divestment of Shutterstock’s entire editorial business, and on 7 July 2026 Getty terminated the agreement, after booking $9.2 million in related costs in the first half of the year. On 22 June 2026, by contrast, it announced a display deal with OpenAI, under which ChatGPT draws on Getty’s licensed catalogue. The value was not disclosed.
This is where the ledger closes counterintuitively. The litigation strategy produced three images and an invoice payable to the opponent; the licensing strategy produced a contract with the largest available buyer. In the United States Getty is still fighting — the claim filed in Delaware in February 2023 was withdrawn and refiled on 14 August 2025 in the Northern District of California, where by mid-2026 the parties were in fact discovery — but the signal the market reads is that the two routes are not moral alternatives. They are two ways of setting a price, and one of them has an open-ended estimate.
One asymmetry no judgment resolves. If a model’s weights are not a copy, the value of a catalogue no longer lies in the individual images but in the right to access them in an orderly way, with metadata, captions and warranties attached. That is precisely the product Getty sold to OpenAI. The London trial did not decide whether that product is protected. It made finding out very expensive.