# Getty v. Stability AI: Why an AI Model Is Not an Infringing Copy

> The English High Court held Stable Diffusion is not an infringing copy for UK secondary infringement: it stores no Getty works and training was abroad.

Topic: Copyright  |  Author: Lidiia Levitska  |  Source: Intellectual Property Law (outsideipcounsel.com)
Canonical: https://outsideipcounsel.com/blog/getty-v-stability-ai-uk-training-location-secondary-infringement/


The first full trial judgment anywhere on whether an artificial-intelligence model itself infringes copyright arrived not from the United States but from London. In *Getty Images (US) Inc. v. Stability AI Ltd*, [2025] EWHC 2863 (Ch), the High Court of England and Wales handed down a landmark decision on November 4, 2025, addressing how UK copyright and trademark law apply to the training and distribution of a generative image model. The court rejected Getty's principal copyright theory, holding that the Stable Diffusion model is not an "infringing copy" for the purposes of UK secondary infringement, because the model stores and reproduces no Getty works, only numerical parameters derived from training.

The ruling is instructive precisely because of what it did not decide. Getty, having accepted that there was no evidence the model's training took place in the United Kingdom, abandoned its training claim, and its outputs and database-right claims fell away too, so it staked its case on secondary infringement premised on importing the model as an infringing article. Territoriality thus became the fulcrum. The decision offers the clearest illustration yet of how the location of AI training and the nature of a trained model determine cross-border copyright liability.

## At a glance

- **Case:** *Getty Images (US) Inc. v. Stability AI Ltd*, [2025] EWHC 2863 (Ch), High Court of England and Wales (Chancery Division).
- **Decided:** November 4, 2025; the first full UK judgment on copyright and generative AI model training.
- **Holding:** Stable Diffusion is not an "infringing copy" for UK secondary-infringement purposes because it does not store or reproduce Getty's works, and with training conceded to have occurred outside the UK, the copyright claims failed; only limited trademark infringement on watermarks was found.
- **Significance:** The leading illustration of how training location and territoriality decide cross-border AI copyright claims, and the first judicial analysis of whether a model is an infringing article.

## The UK secondary-infringement framework

To understand the judgment, one must grasp the structure of UK copyright law under the Copyright, Designs and Patents Act 1988. Primary infringement covers acts like copying a work. It is territorial: the copying must occur in the United Kingdom to be actionable under UK law. Secondary infringement, by contrast, targets commercial dealings in "infringing copies," including importing into the United Kingdom, or possessing or dealing in the course of business, an article that is an infringing copy, under Sections 22 and 23 of the Act.

The secondary-infringement route has a particular logic. An "infringing copy" is generally an article whose making in the United Kingdom would have constituted a copyright infringement, or which has been or is proposed to be imported and whose making abroad would have infringed had it been made here. So the theory Getty pursued asked the court to treat the Stable Diffusion model as an article, imported into the United Kingdom, that was itself an infringing copy of Getty's images. That framing let Getty attempt to reach conduct connected to training that had occurred outside the country.

## The facts and the strategic narrowing

Getty Images, the well-known stock-photography licensor, alleged that Stability AI had scraped millions of Getty images from its websites, without authorization, to train Stable Diffusion, its text-to-image model. Getty initially advanced a suite of claims, including primary copyright infringement based on the training, claims about model outputs, database-right claims, and trademark claims arising from watermarks that appeared in some generated images.

As the case developed, Getty accepted that there was no evidence the training of Stable Diffusion had taken place in the United Kingdom. That concession was fatal to the training claim, because UK primary copyright infringement requires copying within the jurisdiction. Getty abandoned that claim at trial. It also abandoned the outputs claim, on the footing that Stability had blocked the prompts said to have produced the infringing images so that the relief sought had largely been achieved, and with both gone the database-right claim could no longer be advanced. Getty instead relied on secondary infringement: that importing and distributing the trained model in the United Kingdom amounted to dealing in an infringing article. The court's task was to decide whether the model qualified as an "infringing copy," and to resolve the surviving trademark case.

## The court's reasoning: an intangible article that still copies nothing

The judgment made two moves that will be studied closely. First, the court accepted that an "article" for secondary-infringement purposes need not be tangible. Electronic copies stored in intangible media, such as files held in cloud environments, can constitute articles within the meaning of Sections 22 and 23. That holding rejected the argument that only physical objects can be infringing copies and modernized the concept for a digital era, a point favorable to rights holders in principle.

Second, and decisively, the court held that Stable Diffusion is not an infringing copy, because the model does not store or reproduce any of Getty's works. A trained diffusion model consists of learned numerical parameters, weights, rather than stored copies of the training images. Since the model contains no reproduction of Getty's copyrighted photographs, it cannot be an article whose making would have infringed Getty's copyright. The intangible-article ruling therefore did Getty no good: an article that copies nothing is not an infringing copy no matter how it is stored. With the primary claims gone and the secondary claim defeated on this ground, the copyright case failed.

## The narrow trademark victory

Getty did not leave empty-handed, but its win was limited. On the trademark claims, the court described its findings as both historic and extremely limited in scope. Section 10(1) succeeded only as to iStock watermarks generated by v1.x models accessed through DreamStudio or the Developer Platform, on a single example image; the parallel section 10(1) claim on the Getty Images marks was dismissed for want of evidence. Section 10(2) succeeded on those same iStock examples and on a single Getty Images watermark generated by v2.1. The section 10(3) claim, and the claim as it related to SD XL and v1.6, were dismissed, and the judge declined to address passing off. That finding concerned the appearance of Getty-style watermarks in certain outputs, not the underlying copyright question. It was a narrow foothold, confined to specific instances, and did not disturb the central conclusion that the model itself did not infringe Getty's copyrights.

Because Getty did not pursue the primary training, output, and database claims, the judgment does not decide whether UK-based web scraping and model training would infringe. Those territorial and substantive questions were expressly left for future cases or for legislative reform, and the United Kingdom's ongoing policy debate over text-and-data-mining exceptions looms over them.

## Open questions

- **Would UK-based training infringe?** The judgment deliberately does not answer whether scraping and training conducted within the United Kingdom would constitute primary infringement, leaving the core issue for another case or for Parliament.
- **How should courts treat model outputs?** Getty dropped its output-based copyright claims, so the decision does not resolve when a generated image reproduces a substantial part of a training work.
- **What proof would show a model "stores" a work?** The court found the model stored only parameters, not images. Whether different model architectures, or evidence of memorization, could support a finding that a model does reproduce training works remains open.

## Implications for brands and businesses

- **Training location can decide the case.** Because UK primary infringement is territorial, where a model is trained may determine whether copyright claims can be brought at all. Rights holders and developers alike should map where data ingestion and training actually occur.
- **A trained model may not be an "infringing copy."** *Getty* signals that a diffusion model storing only parameters is unlikely to qualify as an infringing article, narrowing the secondary-infringement route against imported models in the United Kingdom.
- **Watermarks and trademarks offer a separate lever.** The limited trademark findings show that when generated outputs reproduce registered marks, trademark law can supply a remedy even where copyright claims fail, so brands should police mark appearance in AI outputs.
- **Watch the policy track.** With the substantive training question unresolved judicially, legislative developments on text-and-data-mining exceptions will heavily influence the UK landscape, and businesses should follow that debate as closely as the case law.

## Frequently asked questions

**What did the English High Court decide in Getty v. Stability AI?**
On November 4, 2025, the court held that Stable Diffusion is not an infringing copy for UK secondary infringement purposes, because the model does not store or reproduce Getty's images, only numerical parameters. With the training conceded to have occurred outside the UK, the copyright claims failed, leaving only limited trademark findings on watermarks.

**Why did the location of training matter so much?**
Getty accepted there was no evidence the training took place in the UK, so it abandoned its training claim and relied on secondary infringement based on importing the model as an infringing article. Because the alleged copying happened abroad, UK primary infringement law did not reach it, making territoriality decisive.

**Did Getty win anything?**
Yes, but narrowly. The court found trademark infringement it called historic and extremely limited in scope, resting on a handful of example images in which Stable Diffusion generated watermarks reproducing the iStock marks, and one reproducing a Getty Images mark. The section 10(3) claim and the broad copyright claims did not succeed.

## Authorities and sources

- [Getty Images v Stability AI, [2025] EWHC 2863 (Ch), official judgment (Courts and Tribunals Judiciary PDF)](https://www.judiciary.uk/wp-content/uploads/2025/11/Getty-Images-v-Stability-AI.pdf)
- [Latham & Watkins, "Getty Images v. Stability AI: English High Court Rejects Secondary Copyright Claim"](https://www.lw.com/en/insights/getty-images-v-stability-ai-english-high-court-rejects-secondary-copyright-claim)
- [Ropes & Gray, "Getty Images Loses Copyright Infringement Claim Against Stability AI in UK's First-of-its-Kind Ruling"](https://www.ropesgray.com/en/insights/viewpoints/102lvxe/getty-image-loses-copyright-infringement-claim-against-stability-ai-in-uks-first)
- [Paul, Weiss, "Getty Images v Stability AI: The UK Courts' First Word on Use of Copyright Works in AI Model Development"](https://www.paulweiss.com/insights/client-memos/getty-images-v-stability-ai-the-uk-courts-first-word-on-use-of-copyright-works-in-ai-model-development)
- [Faegre Drinker, "First Major Judgment on How UK Copyright Law Applies to Artificial Intelligence"](https://www.faegredrinker.com/en/insights/publications/2025/11/first-major-judgment-on-how-uk-copyright-law-applies-to-artificial-intelligence)