Getty v Stability: what the UK court decided
Getty dropped its training and output claims mid-trial and won only a narrow trade mark point. UK training legality was never ruled on.
The England and Wales High Court handed down judgment in Getty's case against Stability AI on 4 November 2025. A lot of the coverage read it as a verdict on whether training an image model on copyrighted photographs is lawful in the UK.
It was not that. Getty dropped the training claim and the output claim partway through the trial, and the court never ruled on either. What survived was a trade mark point the judgment itself described as historic and extremely limited.
If you are making decisions about tools or client work on the basis of "the UK courts have now looked at this," you are working from a case that mostly did not happen.
What actually reached judgment
| Claim | Outcome |
|---|---|
| Training infringement — that copying Getty's images to train the model infringed | Dropped mid-trial. Training was accepted to have taken place outside the UK. Not decided. |
| Output infringement — that generated images reproduced protected works | Dropped mid-trial. Stability blocked the prompts Getty had relied on. Not decided. |
| Secondary infringement — that the model itself was an infringing copy imported into the UK | Failed. The court held the model weights were not a copy of the training images. |
| Trade mark | Narrow win on watermarks in early-version output, characterised as historic and extremely limited. |
Two of the four headline claims never got a ruling. One lost on the merits. One produced a result so narrow the court flagged it as such.
The claim that actually went to judgment on copyright was secondary infringement: that Stable Diffusion was an "infringing copy" imported into, possessed in, or distributed in the UK. The court accepted that an "article" under the CDPA can be intangible. It then held that the model weights were not a copy of Getty's images, because they do not store or reproduce those works. Without a copy, the secondary-dealing provisions did not bite.
That holding is narrower than the headlines made it. It answers whether this model's weights are an infringing copy under those provisions. It does not answer whether the training copies themselves infringed, or whether a generated image can. Those questions were not tried.
Why training never reached the question
UK copyright infringement is territorial. To succeed, a claimant has to establish that the infringing acts happened in the UK. Getty accepted that the training of Stable Diffusion took place outside the UK, so the training claim came out of the case rather than being tested.
That distinction matters for how you read the result. The court did not weigh whether training on copyrighted images without a licence is lawful and conclude that it is. It was never asked the question in a form it could answer. Model training happens wherever the compute sits, and the compute rarely sits where the claimant is.
The practical consequence is that a UK rightsholder suing over training faces an evidential problem before it ever reaches a substantive one. That does not make training lawful in the UK. It makes it unlitigated.
The output claim came out for a different reason. Getty had alleged that generated images reproduced protected works; it dropped that claim after Stability blocked the prompts used to produce the examples. That is a fact about this litigation, not a finding that generated output cannot infringe.
The trade mark win, and what it is actually worth
The surviving finding concerned Getty's marks turning up in generated output — the watermark. The judgment described the infringement as historic and extremely limited: some early versions of the model, some access routes, and a small set of real-world examples, not a continuing or widespread practice. A visible mark reproduced in an image is the easiest thing in this entire area to prove, because you can point at it.
It is not a general rule about generated images, and it should not be read as one. It also does not tell you anything about the harder version of the same question, which is output that reproduces protected expression rather than a protected sign. That is the question the dropped output-copyright claim would have reached.
There is a narrower lesson worth taking, though. If a model reproduces a source watermark, that is a training-data artefact visible in your deliverable, and it is your problem before it is anyone else's. Worth understanding what a watermark is doing in each of the three contexts it shows up in — source artefact, provider-applied mark, and provenance signal — because they get conflated constantly.
What is still open in the UK
The training question has no UK judicial answer. The output-copyright question has none either. There is no domestic precedent either way on either point.
On the statute book, the picture is the status quo plus one untested provision. The government's 18 March 2026 report said a broad text-and-data-mining exception with a rightsholder opt-out was no longer its preferred way forward, which left the existing narrow non-commercial research exception in place. That policy argument is in the UK TDM write-up, and it is a separate question from what this court decided. Section 9(3) of the CDPA, the computer-generated works provision, is still in force. Whether it applies to generative AI output is contested and untested; the same report proposed removing it, but that has not been legislated. Nobody should be relying on it to claim ownership of a generation.
So: no court ruled on training, the secondary claim failed on a copy point, and the only statutory hook anyone points to for owning a generation remains untested. That is the honest position as of August 2026.
How to actually operate in the UK
If you produce commercial content here, the training question is not the one that will cost you money. These are:
- Output. Whether the file you delivered reproduces someone else's protected work is decided on your file. That risk exists regardless of how the training question resolves. The practical version is in our copyright and safety guide for creators.
- Advertising standards. The ASA and CAP position is that there is no blanket duty to disclose AI use. The test is whether the omission misleads — AI imagery implying real product performance, synthetic endorsements, unachievable results. Advertisers are fully responsible for AI-produced content, which means "the model made it" is not a defence.
- The EU, if you sell into it. Article 50 of the AI Act applied from 2 August 2026, and the duties are separate from anything in UK law. We covered what two transparency regimes starting on the same date ask of one producer.
- Provenance discipline. If you are going to claim a file's history, the chain has to survive your edit and export steps, which it usually does not by default. Sign, strip, survive walks the failure points.
On tool selection, the durable filter is not "has this model been sued" but "what does this provider publish about its data, and what does my licence actually permit." That is knowable now. Training-data provenance as a buying criterion is the version for commercial work. When you are comparing options, the model catalog lists what each provider states.
FAQ
Did the UK court decide that training an AI model on copyrighted images is legal?
No. Getty dropped the training claim mid-trial after accepting that training took place outside the UK, so the court never ruled on it. Anyone citing this judgment as UK approval of AI training is citing something that is not in it.
What was the trade mark finding, exactly?
A narrow win concerning Getty's marks — the watermark — appearing in generated output from some early versions of the model. The judgment characterised the infringement as historic and extremely limited. It does not establish a general rule about AI-generated images.
Did the court decide that an AI model is a copy of its training images?
No. On the secondary-infringement claim that did reach judgment, the court held that this model's weights were not an "infringing copy" because they do not store copies of the training images. That is a finding about these weights under the secondary-dealing provisions. It is not a ruling on whether the training copies themselves infringed.
Can I rely on section 9(3) to own an AI-generated image in the UK?
Not safely. The provision exists, but its application to generative AI output is contested and untested in court. If you need a protectable asset, the reliable route is human creative contribution you can evidence — selection, arrangement, editing — rather than a statutory argument nobody has won yet. Treat ownership of the file and copyright in the file as two separate things, and get the usage rights written down either way.