After this ruling, will tech companies move all model training to data centres that they consider “copyright safe”? Will we see a new equivalent of a “tax haven” for training AI models on copyrighted content? An “AI haven”? This article is not legal advice.
Getty Images sued Stability AI in the High Court in London over the alleged unauthorised use of copyrighted images to train Stability AI’s image generation model, Stable Diffusion. A High Court judgment was issued yesterday which probably will have left Getty Images largely disappointed, and which certainly leaves many questions unanswered, and which is likely to have knock-on effects in future for how content creators protect their assets, as well as how tech companies collect training data and train their AI models.[1].
Getty Images is a leading provider of stock images, through its Getty Images brand and also through the website iStockphoto which Getty acquired in 2006. The business involves the licensing of image, video, and audio content. Getty was joined in the litigation by the photographer Thomas M Barwick (to be accurate, I should say, by his company of the same name, which he uses as a trading vehicle), which has 35,000 images and 15,000 videos available on Getty Images’ websites. Mr Barwick was included as an example of a creative content contributor.
Stability AI is a London-based software company which develops generative AI systems including a number of models called ‘Stable Diffusion’, which are well known for generating images from text or image prompts. According to its website, Stability AI counts Titanic director James Cameron among its board of directors, but I could not find a verification of James Cameron being officially listed as a “director” on Companies House so I am unsure what that means in practice.[2, 3]
The judgment included an example of one of Mr Barwick’s stock photos, titled “View from underwater of woman holding balloons”. This has a watermark on it, and I was able to find it on Getty Images’ website where I can buy a digital copy of the image for £275.
The overall aim of the case, from Getty Images’ perspective, was to protect their business’s “lifeblood”, that is, millions of high-quality images, from what they described as exploitation by innovators like Stability AI without regard to the intellectual property rights of creators.
The main points of contention were as follows:
Getty originally claimed that Stability infringed Getty’s copyright by using the Getty content without Getty’s permission and by creating an AI model which could reproduce the copyrighted content.
When you visit Getty’s website, or iStockphotos, before you buy an image, you will see a watermarked image with “GETTY IMAGES” or “ISTOCK”. Getty Images claimed that the fact that Stable Diffusion (or earlier versions of it) have been known to generate images with these watermarks is an “act of trade mark infringement”.
Earlier in the case, Getty had also included a claim for database right infringement, but this was later abandoned.
This case involved an incredible amount of technical detail which would have been unfamiliar to the judge and lawyers involved. For that reason, both sides included evidence from expert witnesses as well as witnesses of fact. The witnesses of fact included people with knowledge of the matter such as current and former employees, and the expert witnesses were two professors in the field of AI models.
It was interesting to read the summary of the expert witnesses’ evidence, which mostly consisted of a joint statement. In England and Wales, an expert witness is more tightly constrained than in many other jurisdictions, such as the USA. A witness of fact provides evidence in court about events that they were personally involved in but cannot offer personal opinions. An expert witness’s primary duty is to assist the court in matters within their expertise, by providing objective and unbiased opinions. Expert witnesses in civil litigation in England and Wales (such as the Getty vs Stability case) have to adhere to the Civil Procedure Rules[7], and other jurisdictions in the UK have equivalent sets of rules. Crucially, these rules were introduced to avoid expert witnesses becoming ‘hired guns’ for either side in litigation.
There is also an emphasis in the Civil Procedure Rules on encouraging both sides in civil litigation to jointly instruct an expert, so I would be interested to find out why that did not occur in this case, and both sides instructed their separate experts who ended up agreeing on most things and writing a joint statement.
The judge also commended the two experts for producing a helpful “Agreed Technical Primer” which provided a summary of the technology relevant to the case.
Experts – overriding duty to the court
(1) It is the duty of experts to help the court on matters within their expertise.
(2) This duty overrides any obligation to the person from whom experts have received instructions or by whom they are paid.
- Civil Procedure Rules 35.3
Fast Data Science - London
A large part of the legal questions addressed by the judge hinged on nuances that are not related to questions of technology, for example, whether Stability AI had done any of the training of the model in the UK (both sides agreed that it had not done so), and also to what degree the development of the Stable Diffusion model was the responsibility of Stability AI and to what degree it was responsibility of a group called CompVis, the Computer Vision and Learning research group at a Germany university, LMU Munich. Furthermore, it was found that only historic versions of Stable Diffusion generated fake Getty Images watermarks.
It is a common source of confusion for laypeople, whether models such as ChatGPT have “memorised the entire internet” or have just learnt patterns from the data that they have been trained on. In fact, an AI model trained on large amounts of data learns patterns from that data but does not store the original data.
In their Agreed Technical Primer, the expert witnesses wrote that models such as Stable Diffusion can be prone to memorisation, that is, reproducing images used in training. They explain that to work reasonably, a model needs to be able to ‘generalise’, that is, apply patterns in a different context. A model which has been trained too long on the same data or trained on insufficiently diverse training data is prone to ‘overfitting’, which is an undesirable outcome.
Interestingly, the experts cited a study (“the Carlini study”[5]) which found that a diffusion model such as Stable Diffusion could memorise an image from the training set if that image is reproduced between 200 and 3000 times, although what is unclear is if many different images contain the same watermark, how many images would be needed to reproduce the watermark.
Despite this, Mrs Justice Smith wrote in her judgment that
an AI model such as Stable Diffusion which does not store or reproduce any Copyright Works (and has never done so) is not an “infringing copy”
It is worth pointing out that Getty Images offers its own “commercially safe” paid image generation model, which has been trained “exclusively on licensed visuals” (Generative AI by Getty Images). Getty also has a partnership with Perplexity, another AI firm, which allows Perplexity to display images from Getty across its AI-powered search engine tools.[8]
Last year I wrote a blog post discussing whether generative AI is a threat to creative industries, reporting that traditional content providers such as stock photo libraries were beginning to offer their own generative AI models, which can be guaranteed safe for commercial use, with no celebrities, brand names, or other protected content. I am not sure how successful these ventures have been.
Getty’s claim was largely unsuccessful. The copyright claim was rejected by the court, but Getty won the trademark claim with a very limited scope (historic versions of Stable Diffusion did generate Getty watermarks).
The dismissal of the copyright infringement claim came down to the court’s interpretation of UK copyright law in the context of AI models. However, Getty voluntarily abandoned some claims prior to the final ruling (the claim that the model’s output infringed on their copyright, the claim that the training and development infringed on copyright, and the database rights infringements claim).
Since it was agreed by both sides that Stable Diffusion was trained outside the UK, Getty was unable to pursue the claim that the model training itself violated their copyright. Stability AI argued that all the computing resources used for training were located in two U.S. data centers operated by Amazon Web Services (AWS).
Yesterday’s High Court ruling has been described as a ’landmark case’ which sets a precedent for the use of copyrighted material in model training.
The dismissal of the copyright claim on the grounds that the AI model does not store or reproduce any copyright works has been described by lawyers as exposing weaknesses in Britain’s copyright protections.[6]
As I understand it, this interpretation would mean that I would be committing a copyright infringement if I were to post all seven Harry Potter books online, but if I were to train a machine learning model on the same content and post that model online, I would not have committed an infringement, on the grounds that my model has learned only patterns but does not store the copyrighted content.
The judgment makes it clear that “whether an article is tangible or intangible, an infringing copy must be a copy; i.e. a reproduction of a Copyright Work.”, and since “the model weights of the various Stable Diffusion versions do not store the visual information in the Copyright Works, they cannot amount to a copy.” I guess part of this boils down to the training set and if a model were trained and sufficiently overfit on a training dataset, there would be a risk of copyright infringement.
In practical terms, it was interesting to me to notice that the judgment factored in issues such as which organisation’s Github account and HuggingFace account certain models were hosted on. For example, I have hosted models and code in different Github accounts without considering that this could be something that is cited in a court judgment about which organisation was responsible for a particular open source software product or machine learning model.
It is also instructive to note that if training a model on copyrighted data did constitute any kind of copyright infringement, that is dependent on the jurisdiction that the training takes place, in this case the location of the AWS cluster that Stability AI used. So in the future, AI companies may move all model training to AI-friendly jurisdictions as opposed to jurisdictions with strong copyright protections - perhaps similar to how some countries with favourable tax regimes can be “tax havens” for multinational tech companies.
The case will continue, as Getty vs Stability is only the first of the copyright cases as new technology has opened up uncharted legal territory. Getty is continuing with litigation in other jurisdictions.
Getty Images vs Stability AI, [2025] EWHC 2863 (Ch), accessed at https://www.judiciary.uk/wp-content/uploads/2025/11/Getty-Images-v-Stability-AI.pdf, accessed 5 Nov 2025.
Stability AI Ltd, Companies House, https://find-and-update.company-information.service.gov.uk/company/12295325, accessed 5 Nov 2025.
https://stability.ai/news/james-cameron-joins-stability-ai-board-of-directors, accessed 5 Nov 2025.
Civil Justice Council, Guidance for the instruction of experts in civil claims, https://www.judiciary.uk/wp-content/uploads/2014/08/experts-guidance-cjc-aug-2014-amended-dec-8.pdf, accessed 5 Nov 2025.
Carlini, Nicolas, et al. “Extracting training data from diffusion models.” 32nd USENIX security symposium (USENIX Security 23). 2023.
Sam Tobin, Getty Images largely loses landmark UK lawsuit over AI image generator, Reuters, https://www.reuters.com/sustainability/boards-policy-regulation/getty-images-largely-loses-landmark-uk-lawsuit-over-ai-image-generator-2025-11-04/, accessed 5 Nov 2025.
Civil Procedure Rules, https://www.justice.gov.uk/courts/procedure-rules/civili
Getty Images press release titled Getty Images and Perplexity strike multi-year image partnership, https://investors.gettyimages.com/news-releases/news-release-details/getty-images-and-perplexity-strike-multi-year-image-partnership, accessed 5 Nov 2025.
Dive into the world of Natural Language Processing! Explore cutting-edge NLP roles that match your skills and passions.
Explore NLP Jobs
This new video explains natural language processing: what it is, how it works, and what can it do for your organisation. Natural Language Processing (NLP) is a branch of Artificial Intelligence (AI) that focuses on giving computers the ability to understand human language, combining disciplines like linguistics, computer science, and engineering.
This free A/B test calculator will help you compare two variants of your website, A and B, and tell you the probability that B is better. You can read more about A/B testing in our earlier blog post on the subject. You may also be interested in our Chi-Squared sample size calculator which will help you calculate the minimum sample size needed to run a Chi-Squared test, given an expected standardised effect size.
See also: Fast Data Science A/B test Calculator (Bayesian) A/B testing is a way you can test two things, Thing A, and Thing B, to see which is better. You most commonly hear about A/B testing in the context of commercial websites, but A/B testing can be done in a number of different contexts, including offline marketing, and testing prices.
What we can do for you
