Getty Images vs. Stability AI: Copyright Explained

A quick, but incisive, guide to what the copyright case is about, why Getty Images lost, and the findings being appealed. 

What’s the case about? 

The dispute concerns Stable Diffusion, a generative AI tool created by the defendant, Stability AI. Stable Diffusion is an image generation tool—users enter a text or image prompt, and the tool creates an AI-generated (synthetic) image in response. Stable Diffusion was trained on millions of images taken from the internet, including Getty Image’s vast image library. Getty therefore brought claims against Stability for copyright, database right, and trademark infringement. Here, we cover copyright only. 

What were the copyright claims?  

Copyright infringement falls into two broad categories: primary infringement (which includes copying a work and making copies of a work available to the public) and secondary infringement (which includes dealing with infringing copies). Getty alleged both primary and secondary infringement against Stability.

Copyright, like all IP rights, is territorial in nature, so infringing activities must occur in the U.K. for a finding of infringement here. Getty has brought a separate action against Stability in the U.S., where most of the training and development of Stable Diffusion took place. 

Why were the primary copyright infringement claims dropped? 

Getty advanced two claims of primary copyright infringement, both of which were dropped at the end of the trial. 

• Infringement by copying during training/development. Stability did not deny that it had copied the images in Getty’s image library during the training/development of Stable Diffusion. Accordingly, the court did not have to decide whether there was copying in the sense required by section 17 or whether the temporary copies exception applies. Rather, Stability argued no infringement on the basis that no copying took place in the U.K. While most training/development clearly took place in the U.S., Getty pursued the U.K. claim because some of Stability’s employees/contractors were based in the U.K., so it was quite possible that infringing activities took place here. Getty did its best to evidence a U.K. infringement, but the complex way that Stable Diffusion was trained made it difficult to piece together what happened where. This was exacerbated by Stability’s limited record-keeping and witness evidence. Getty did not help itself either—in the rush to trial, it failed to follow up new lines of enquiry arising from Stability’s disclosure in the correct way, so any additional evidence was not flushed out, and Getty could not ask the judge to draw adverse inferences arising from gaps in Stability’s disclosure/evidence. The final ‘nail in the coffin’ of Getty’s claim came at trial when Stability’s witnesses gave fairly convincing evidence that all training/development of Stable Diffusion was undertaken on non-U.K. cloud-based servers and that it would not have been necessary or desirable to download any training datasets onto U.K. servers or computers. This left Getty little option but to drop this claim. It does not mean that all such claims of copying will fail, however.

• Infringement by communication to the public/authorising acts of copying at the output stage. Getty also claimed that Stable Diffusion infringes at the ‘output’ stage under section 20 and/or 16(2)/17, by producing synthetic images that copy a substantial part of works on which the model was trained. An example of an input image and allegedly infringing output image are at the end of this article. Getty says that it dropped this claim at the end of the trial because Stability agreed to block all use of the prompts complained of, negating any need for an injunction. While we don’t know exactly what was agreed between the parties, the blocking of a limited number of prompts does not seem like much of a victory for Getty. More likely, this claim was primarily dropped/settled because the allegedly infringing outputs weren’t similar enough to any Getty image to show copying of a substantial part, especially given the arguably low level of originality of the Getty images relied upon. Not only that, but Getty had numerous other hurdles to overcome to succeed on this claim, including proving title to the Getty images alleged to be infringed, that infringement was likely in the real world (and not just a result of the use of contrived prompts), and various legal issues such as that there can be a “communication to the public” if only one user receives the synthetic image output. While it is disappointing that we won’t receive a ruling on these novel issues, there is still learning from this case. It is clear that it will take a different type of case for a finding of “output” infringement, perhaps one where the “input” image is highly original/distinctive and the “output” image is very close to it. 

Why did the secondary copyright infringement claim fail?  

With the claims of primary infringement having been dropped, Getty’s claim of secondary infringement was the only one on which the court ruled. The secondary infringement claim is that Stability has infringed by importing and/or distributing in the U.K. an ‘article’ (namely, the Stable Diffusion model) which is—and which Stability knew or had reason to believe is—an ‘infringing copy’ of Getty’s copyright works (under sections 22 and 23).

This claim largely hung on two questions: whether an ‘article’ can be something intangible like an AI model and whether the Stable Diffusion model meets the definition of an ‘infringing copy’. While the court found for Getty on the first issue, it did not on the second, meaning that Getty's claim failed overall.

• Can an ‘article’ be something intangible? The court concluded that an ‘article’ can be something intangible on the basis that, since the relevant legislation was enacted,storage has developed from tangible items (such as floppy disks) to intangible items (such as the cloud), and Parliament would not have wanted to limit the legislation to the former. While there is some force to this argument, it fails under scrutiny. There is, of course, no such thing as intangible storage. The cloud is actually just a tangible server located somewhere, and any import/distribution of information stored in the “cloud” is just an electrical transmission. The import/distribution of an AI model this way is therefore more akin to a “communication to the public” (primary infringement) than a dealing in an “article” that is an “infringing copy” (secondary infringement). Mixing these regimes is questionable—they are very distinct, have different knowledge requirements, and were carefully considered by Parliament when enacted.

• Is Stable Diffusion an ‘infringing copy’? The court held that Stable Diffusion does not meet the definition of an ‘infringing copy’, although this part of the ruling has been substantially misinterpreted by some commentators.Importantly, the court did not have to answer whether Stable Diffusion is an ‘infringing copy’ because it stores copies of the Getty works on which it was trained. Getty did not run this argument (and so the case is quite different to the German GEMA case). Rather, Getty argued that Stable Diffusion is an ‘infringing copy' even if it does not store copies of Getty’s works because section 27 provides that an ‘article’ is an ‘infringing’ copy if “…its making in the U.K. would have constituted an infringement of the copyright in the work in question…” In other words, Getty ran a ‘but for’ argument—‘but for’ the Getty images existing in the training dataset, the Stable Diffusion model could not have been made. Ultimately, this argument failed. The court held that the secondary infringement regime is concerned with downstream dealings in an ‘article’ that is an ‘infringing copy’, not with a process which (while it may involve acts of infringement) ultimately produces an ‘article’ which is not itself an ‘infringing copy'. There is some force to this argument despite the wording of section 27. Furthermore, a finding that an ‘infringing copy’ can be something that is not itself actually a copy could mean that there is secondary infringement where an ‘article’ is produced with the benefit (anywhere in the process) of an infringing work, subject to the knowledge and other requirements. This would be a significant change to the law, with numerous consequences. 

What about the appeal? 

Getty has been granted permission to appeal the finding on ‘infringing copy’ to the Court of Appeal. If Stability asks it to do so (which is likely), the appeal court will also have to consider whether an ‘article’ can be something intangible like an AI model, meaning the court will have to consider both of the above questions.

It is difficult to predict what the Court of Appeal will rule as the legislation is unclear and the issues can be argued either way. However, if the Court of Appeal does find secondary infringement, there is a separate and intriguing question for the parties to address—whether only the downloadable version of Stable Diffusion infringes or whether the (more commonly used) web-based version also infringes. Given that the web-based version is held on servers based outside of the U.K., the latter seems a difficult argument. This would seem to give Stability a workaround even if there is a finding of secondary infringement.

The appeal has recently been passed to the listing office, but is unlikely to be heard until the end of 2026 at the earliest, perhaps with a decision in spring/summer 2027.