From Ritu Raj | Product & Market Analysis
Getty v Stability AI: Why the UK Court Ruled Model Weights Are Not Copies
On this page
Getty Images lost the copyright half of its UK case against Stability AI on a single sentence of statutory reasoning. At paragraph 600 of a judgment handed down on 4 November 2025, Mrs Justice Joanna Smith held that Stable Diffusion's model weights "have never contained or stored an infringing copy" of Getty's photographs. That makes the Getty v Stability AI ruling the first UK answer to whether a trained model is itself a copy. The answer was no, and it is now on appeal.
Key takeaways
- The UK High Court held that Stable Diffusion's weights are not an infringing copy of the images it was trained on. The judge found the weights store learned patterns, not the works, so importing and distributing the model was not secondary infringement.
- The biggest questions in the case were never decided. Getty abandoned its training claim and its output claim before closing submissions, partly because there was no evidence that training took place in the United Kingdom.
- The ruling protects whoever imports, holds or distributes a model in the UK, which is exactly where a deployer sits. Had Getty won, possessing downloaded weights in the course of business could have become an infringing act.
- Other courts are pulling in different directions. A Munich court found song lyrics memorised by ChatGPT were copies, and the US Third Circuit held in September 2026 that training a non-generative tool on Westlaw headnotes was not fair use.
What the Getty v Stability AI ruling actually decided
The case is Getty Images (US) Inc and others v Stability AI Ltd, neutral citation [2025] EWHC 2863 (Ch). It was tried over June 2025 in the Intellectual Property List of the High Court and decided on 4 November 2025. Most coverage summarised it as "Stability wins". That is true and it is also the least useful way to read it.
The judgment decided three things that matter to anyone building on AI models. First, an "article" under the UK copyright statute can be intangible, so a file sitting in cloud storage can in principle be an infringing copy. Second, Stable Diffusion's weights are not such a copy, because they never stored Getty's works. Third, Stability infringed Getty's trade marks in a narrow set of cases where older model versions generated Getty and iStock watermarks.
Everything else was either abandoned or left alone. If you read one paragraph of the judgment, read paragraph 9, which records what Getty gave up. If you read two, add paragraph 600, which is the holding on model weights.
Our view is that the ruling is more significant for deployers than for model developers. Developers trained outside the UK and were never going to be caught by this claim. Deployers import, download and hold models here, and that is the conduct secondary infringement law targets.
The claims Getty dropped before the judge ruled
Getty started with a much larger case. It alleged that Stability scraped its images to train Stable Diffusion, that outputs reproduced its works, and that its database right was infringed. The letter of claim of 16 January 2023 said Getty content "said to be 12 million links" had been scraped.
Shortly before closing submissions, the case shrank. Paragraph 9 records three concessions. Getty accepted that "there is no evidence that the training and development of Stable Diffusion took place in the United Kingdom". It treated the outputs claim as "substantially achieved" because Stability had blocked the prompts behind its example outputs. And the database right claim "can now no longer be advanced".
The training point is jurisdictional, not philosophical. UK copyright is territorial. Paragraph 32 of the judgment records that the compute cluster Stability used was located outside the United Kingdom. A claim that training infringes UK copyright needs training to happen in the UK, and Getty could not prove it did.
So the headline question, whether training a model on scraped images is lawful, was not answered. Anyone telling you the UK has ruled that AI training is legal has misread the case. The court ruled on what happens when a finished model crosses the border.
Why the court held model weights are not an infringing copy
Getty's surviving copyright case was secondary infringement under sections 22 and 23 of the Copyright, Designs and Patents Act 1988. Those sections make it unlawful to import, possess in the course of business, or distribute an "article" that you know or have reason to believe is an "infringing copy". The argument was that Stable Diffusion, imported into the UK through downloads, was that article.
Getty did not claim the model contained its photographs. Paragraph 11 records that Getty "do not say that Stable Diffusion is itself a copy of, or that it stores within it any copies". Its argument was subtler. Under section 27(3), an article is an infringing copy if making it in the UK would have infringed. Getty said making the weights required copying its images, so the weights were infringing copies.
An article can be intangible
Stability's first defence was that an article has to be a physical thing. The judge disagreed. At paragraph 583 she held that "an electronic copy stored in an intangible medium (such as the AWS Cloud) is, in my judgment, capable of being an infringing copy and thus also capable of being 'an article'."
That holding is a loss for AI developers, and it gets less attention than it should. It means a file downloaded from a model hub can, in principle, be imported into the UK in the legal sense. The court also found that Stability's downloadable weights had been taken up in the UK in volume, with "many downloads ... (perhaps running into millions)".
But a copy has to be stored at some point
The case turned on the next step. At paragraph 597 the judge wrote: "I cannot see how an article can be an infringing copy if it has never consisted of/stored/contained a copy." At paragraph 600 she applied that to the evidence. The weights "are purely the product of the patterns and features which they have learnt over time". In its final form, Stable Diffusion "does not store or reproduce any Copyright Works and nor has it ever done so."
The judge treated the copying during training as irrelevant to the status of the end product. Exposure of the weights to copies of Getty's works during training did not turn the result into a copy. We think that is the correct reading of the statute as drafted, and also the reading most exposed on appeal, because it separates a product from the process that made it.
| Question | Getty's position | What the court held |
|---|---|---|
| Can an intangible file be an "article"? | Yes, including cloud-stored weights | Yes. Getty won this point (para 583) |
| Must an infringing copy contain the work? | No. It is enough that making it would infringe | Yes. It must have consisted of, stored or contained a copy (para 597) |
| Do Stable Diffusion's weights contain Getty works? | Not argued | No, and they never did (para 600) |
| Does a hosted service import a copy? | Yes, through DreamStudio and the API | No. "No copy of the Model is ever provided to the user" (para 604(iii)) |
| Where Getty's case was stronger | Knowledge | The judge would have found Stability knew or had reason to believe, had the weights been copies |
Paragraph numbers refer to the judgment text published by The National Archives. The knowledge finding sits in the judge's alternative analysis, written in case she was wrong on the main point.
The trade mark win Getty kept
Getty did win something. Its photographs carry visible Getty Images and iStock watermarks, and early versions of Stable Diffusion sometimes generated images bearing distorted copies of them. The court found trade mark infringement under sections 10(1) and 10(2) of the Trade Marks Act 1994 for a limited set of those outputs.
The limits are the story. The findings were confined to older model versions in the v1.x and v2.x families, and tied to specific prompt types such as "news photo" and "vector art". At paragraph 218(v) the judge recorded there was no evidence in respect of v1.6 or SD XL. The court reached its conclusions on specific examples, not on a general finding that the model reproduces marks.
For a deployer, the lesson is practical rather than doctrinal. The liability Stability actually incurred came from outputs, not training. A model that regurgitates a recognisable mark is a problem whatever the copyright position, and later versions largely fixed it. That is a test you can run yourself.
How the UK ruling compares with US AI copyright lawsuits
US law asks a different question. The live US AI copyright lawsuit is usually about fair use: whether copying works into a training set is permitted because the use is transformative and does not harm the market for the originals. The UK has no general fair use defence, which is one reason Getty's UK case was built on secondary infringement instead.
Fair use, not secondary infringement
Two district judges in San Francisco ruled on training in June 2025. In Bartz v Anthropic, the court found training on lawfully acquired books was fair use but declined to extend that to books downloaded from pirate libraries. In Kadrey v Meta, the court granted Meta summary judgment on fair use, while saying the plaintiffs had failed to evidence market dilution, as White & Case summarised both rulings.
The pirate library point is what cost Anthropic money. The $1.5 billion settlement received final approval on 20 July 2026, at roughly $3,000 per work. What that payment did to the price of training data is covered in our breakdown of the $1.5 billion repricing of training data, so we will not repeat it here.
The first appellate ruling went the other way
On 29 September 2026 the Third Circuit decided Thomson Reuters v Ross Intelligence. It affirmed that Ross's use of Westlaw headnotes to train a legal research tool was not fair use. Judge Tamika Montgomery-Reeves called it "no more than an ordinary copyright case".
The court stressed that Ross's tool "was not a generative AI", according to Mitchell Silberberg & Knupp's analysis. It also found Ross harmed "the market for licensing headnotes as text to train AI", which it called "rapidly developing". That second point matters more than the first. Once a court accepts that a training-data licensing market exists, unlicensed training starts to look like lost licence revenue.
Getty's own US case is still alive. After dismissing its Delaware suit, Getty refiled in the Northern District of California. In a ruling summarised by Loeb & Loeb in April 2026, the court denied Stability's motion to dismiss except on a false copyright management information claim. The copyright infringement claim was not part of that motion.
Where Munich disagreed: memorisation inside the model
A week after the UK judgment, the Regional Court in Munich ruled for the German collecting society GEMA against OpenAI. According to GEMA's account of the 11 November 2025 ruling, the court found that "the ChatGPT systems contain copies of the original works", meaning song lyrics, which were reproduced in response to simple prompts.
The two judgments are less contradictory than they look. The UK court was answering a statutory question on evidence that Stable Diffusion did not store Getty's images. The Munich court, as GEMA reports it, was dealing with a model that could output lyrics verbatim. Memorisation is a fact about a specific model and a specific work, not a property of all neural networks.
That is the useful frame for anyone deploying models. The UK holding protects you to the extent your model behaves like Stable Diffusion did on the evidence. If your model reproduces protected text or images word for word, the Getty reasoning gives you much less cover. GEMA's description is a rights-holder's characterisation, and we have not read the Munich judgment itself, so treat the detail as provisional.
What model weights copyright means for deployers
If you run a product on top of a third-party model in the UK, sections 22 and 23 are the provisions that would have reached you. They cover importing, possessing in the course of business and distributing infringing copies. A company that downloads open weights onto its own servers does at least the first two.
The judgment removes that risk for now, on these facts. It does not remove it permanently. If the Court of Appeal reverses paragraph 600, the analysis flips, and every UK business holding weights trained on unlicensed data would have to think about the knowledge element in section 23.
Downloaded weights versus hosted APIs
The most practical finding in the judgment is about hosting. At paragraph 604(iii) the court held that DreamStudio, Stability's hosted service, never provided a copy of the model to the user. The secondary infringement claim against the hosted service failed for that reason alone, separately from the weights point.
So the architecture you pick changes your legal position. Calling a model through an API means you never import or possess the weights. Self-hosting open weights means you do. That does not make self-hosting unwise, but it should be a deliberate choice, and the licence terms that govern it are compared in our review of open-weight model licence clauses.
Outputs are still open
Nothing in the UK judgment says outputs are safe. The outputs claim was abandoned, not decided. The trade mark finding shows the one place where Stability actually incurred liability was in what the model generated. If your product publishes generated images or text to customers, output review is the control that matters.
Contract terms are the second control. Many providers now offer output indemnities with exclusions you need to read, and the differences between them are set out in our comparison of model provider terms and in the breakdown of coding assistant indemnities. An indemnity is only as good as its carve-outs.
| How you use the model | Secondary infringement exposure after Getty | What still bites |
|---|---|---|
| Hosted API from a provider | Low. No copy is provided to you (para 604(iii) reasoning) | Outputs you publish, provider terms |
| Self-hosted open weights | Low while para 600 stands. Rises if the appeal succeeds | Knowledge under section 23, licence terms, outputs |
| Fine-tuning on your own customers' content | Not addressed by this case | Your own data rights and contracts |
| Distributing a model to your customers | Low while para 600 stands | Section 23(d) distribution, if the appeal reverses |
The EU adds a separate layer. Downstream obligations under the general-purpose AI rules do not depend on how the Getty appeal comes out, and they are mapped in our guide to EU GPAI enforcement for downstream deployers.
Where this reading of the judgment is weakest
We have argued the ruling is good news for UK deployers with a short shelf life. Here are the reasons that argument could be wrong, in both directions.
It is a first-instance judgment under appeal
The High Court granted Getty permission to appeal the secondary infringement finding in December 2025. According to Wiggin's note on the permission decision, the judge said the issue "concerns a pure question of law, namely a matter of statutory construction". Trial judges do not usually describe a point that way unless they think a higher court might reasonably decide it differently.
We found no reported Court of Appeal hearing date in the sources we checked as of 19 October 2026. Stability may also challenge the finding that an article can be intangible. If it wins that argument, deployers end up safer than they are today. If Getty wins on paragraph 600, they end up less safe.
The finding is about one model's facts
The holding rests on evidence about how Stable Diffusion's weights work. A different model, trained differently, with measurable memorisation of specific works, could produce a different finding on the same statute. The Munich outcome shows a court willing to find copies inside a model when the outputs prove they are there.
There is also a policy risk the judgment cannot settle. The UK government's March 2026 report said a broad exception with an opt-out "is no longer the government's preferred way forward", and that it would not reform copyright law until confident of its objectives. That leaves courts filling the gap, case by case. The weakest assumption in any deployer plan built on this ruling is that the gap stays where Getty left it.
Frequently asked questions
What did the court decide in Getty v Stability AI?
The UK High Court ruled on 4 November 2025 that Stable Diffusion's model weights are not an infringing copy of Getty's images, because they never stored those images. Getty's secondary copyright claim therefore failed. Getty won a narrow trade mark claim over watermarks generated by older model versions. Its training, output and database right claims were abandoned before the judge ruled on them.
Are AI model weights a copy of training data under UK law?
On the evidence in Getty v Stability AI, no. The High Court held that an infringing copy must at some point consist of, store or contain a copy of the work. It found Stable Diffusion's weights contain learned patterns, not the images. The ruling is first-instance, it is under appeal, and a model shown to memorise specific works could be treated differently.
Did the UK ruling say AI training on copyrighted data is legal?
No. Getty abandoned its training claim because there was no evidence that Stable Diffusion was trained in the United Kingdom, and UK copyright is territorial. The court never decided whether training a model on scraped images in the UK infringes copyright. That question remains open, and the UK government said in March 2026 it would not change the law for now.
Is Getty appealing the Stability AI judgment?
Yes. In December 2025 the trial judge granted Getty permission to appeal the secondary infringement finding, describing it as a pure question of statutory construction. Stability was refused permission by the trial judge to appeal the trade mark finding. We found no reported Court of Appeal hearing date as of 19 October 2026 in the sources we checked.
How is the UK Getty ruling different from US AI copyright lawsuits?
US cases mostly turn on fair use, a defence the UK does not have in that form. US district courts found training on lawfully acquired books was fair use in 2025, while Anthropic paid $1.5 billion over pirated books. In September 2026 the Third Circuit held that training a non-generative legal research tool on Westlaw headnotes was not fair use.
Does the Getty ruling protect companies that deploy AI models in the UK?
Partly, and for now. Secondary infringement covers importing, possessing and distributing infringing copies, which is what deployers do with downloaded weights. The ruling says weights are not copies, so that route is closed while the judgment stands. It says nothing about outputs, and hosted APIs carry less exposure because no copy of the model reaches you.
Where to start this week
Make one list. For every model your company uses in production, write down whether you call it through a hosted API or hold the weights yourself, and which version you run. The judgment makes that distinction legally meaningful, and most teams cannot answer it from memory.
Then run a cheap regurgitation test on the self-hosted ones. Prompt for well-known protected text, logos and watermarks, and keep the results with a date on them. If the appeal goes against Stability, that file is the evidence of what you knew and when. If you want a format for recording model provenance, our AI bill of materials template already has the fields.
Related on this site
For the money side of the same disputes, read the tally of AI content licensing deals, which shows what rights holders are being paid outside court.
References
- High Court of England and Wales, Getty Images (US) Inc and others v Stability AI Ltd [2025] EWHC 2863 (Ch), 4 November 2025. Used for all paragraph-referenced findings, quotes and dataset figures.
- Courts and Tribunals Judiciary, Getty Images v Stability AI judgment page, November 2025. Used for the date, judge and citation.
- UK Government, Report on Copyright and Artificial Intelligence, 18 March 2026. Used for consultation response count and policy position.
- Wiggin LLP, High Court grants permission to appeal in Getty Images v Stability AI, 19 January 2026. Used for the permission decision and quote.
- Loeb & Loeb, Getty Images (US), Inc. v. Stability AI, Ltd., 23 April 2026. Used for the US motion to dismiss outcome.
- Authors Guild, Court grants final approval of Anthropic copyright settlement, 21 July 2026. Used for the Bartz settlement figures.
- Mitchell Silberberg & Knupp, Third Circuit: training an AI tool on Westlaw headnotes is not fair use, 1 October 2026, with LawNext coverage, 12 October 2026. Used for the Ross appellate ruling.
- GEMA, Landmark ruling: GEMA against OpenAI, 11 November 2025. Used for the Munich outcome, as characterised by the claimant.
Weakest point in the source base: the Munich and US appellate rulings are reported here through a party's press release and law firm summaries, not the judgments themselves. The UK findings come from the judgment text. This post is analysis, not legal advice.
Related reading