This has become more apparent following Getty’s withdrawal during the trial’s closing submissions of its primary copyright infringement allegations regarding training and output. This raises the general question of how national copyright laws apply to the training and deployment of AI models around the world.
Getty Images et al. v. Stability AI: background
Getty Images (Getty) filed an action in the UK (and an equivalent in the U.S.) alleging copyright, trademark, and database rights infringement. In the UK action, Getty originally asserted that Stability AI unlawfully copied in the UK millions of images, protected by UK copyright and owned or represented by Getty Images, to train its Stable Diffusion image generator.
It also claimed that there was further unauthorized copying or communication to the public in the UK of a substantial part of its images at the point of use, i.e., in the images output from Stable Diffusion. These claims were, however, withdrawn by Getty on the first day of the trial’s closing submissions.
The importance of where to train AI systems
Stability AI had admitted that at least some Getty images were used to train Stable Diffusion. However, Getty’s difficulty with its training allegation was proving that any infringing act occurred in the UK, i.e., whether Getty’s works were copied in the UK during training such as by downloading them onto hardware in the UK.
Stability AI had argued that no infringing acts took place in the UK in training Stable Diffusion because Stable Diffusion was wholly developed and trained outside of the UK and, as such, there was no infringement of UK copyright in this respect.
As this claim has now been withdrawn, we are unlikely to get a judgment on that issue. Nevertheless, it is of real importance to any company deciding where to develop and deploy its AI models. Copyright protection is territorial, and at a high level, the jurisdiction where any infringing acts take place affects the risks of copyright infringement, in particular which defenses to copyright infringement are available.
Currently, different countries have differing approaches to exempting copyright reproduction for text and data mining (TDM) purposes:
- Japan and Singapore allow TDM for commercial purposes.
- The EU allows copyright owners to opt-out works from commercial TDM.
- The UK currently only allows TDM for non-commercial research. However, as explained in our previous blog post, it is considering introducing an exception for commercial TDM with the ability for rights holders to reserve their rights (an “opt-out”).
The U.S. takes an entirely different approach, deciding on the particular facts of each case whether the use in training could be considered to be “fair.” In one case, the use of legal headnotes to train a competitor’s AI tool was held not to be fair use given the particular use case1. On the other hand, in a recent federal court judgment in California, AI foundation model developer Anthropic has been allowed to assert fair use against copyright claims for training its Claude AI models on copyrighted books that Anthropic had lawfully acquired2.
However, the judge in that case ruled that the same “fair use” argument would not apply in respect of Anthropic’s collection and use of pirated works, which will be the subject of a separate damages hearing (with potentially significant amounts at issue).
Meanwhile, in a separate federal court proceeding in California against Meta, the judge concluded that “fair use” was likely not available given the market dilution impact of large language models (although the judge found in favor of Meta for other reasons)3.
Accordingly, reproducing copyright works during AI training may be exempted in countries with more expansive exceptions, but it will only be possible to fully take advantage of such expansive exemptions if one can restrict all acts of reproduction to that country and nothing is downloaded or stored elsewhere.
Furthermore, these TDM exemptions will only apply (if at all) to the steps of training the AI model. Subsequent acts of reproduction or communication to the public following completion of the AI training may involve new infringing acts that may not be covered by the TDM exceptions. These may include, for example, further reproduction (e.g., uploading protected content) or making any data resulting from TDM activities available to the public (e.g., making it accessible on the internet).
Deployment
Despite the withdrawal of the training claim, Stability AI will not necessarily avoid UK copyright liability. This is because Getty continues with two secondary infringement claims that cover the deployment of Stable Diffusion in the UK:
- Stable Diffusion is alleged to be an article that is, and that Stability AI knows or has reason to believe, is an infringing copy of Getty’s copyright works, which has been imported into the UK (s23 CDPA ’88).
- Stability AI is alleged to have possessed and distributed in the course of business, sold, offered, or exposed for sale, an article that is, and that it knows or has reason to believe, is an infringing copy (s33 CDPA ’88).
These allegations raise some interesting legal questions:
Can an intangible AI model like Stable Diffusion that is accessible from the internet ever constitute an “article” capable of being imported or dealt with in the course of business?
Can the AI model itself (like Stable Diffusion) be said to be an infringing copy (i.e., a substantial reproduction) of one or more works that were used to train the model?
What if Stable Diffusion does not retain internally in the model any copies of any of the works on which it was trained in material form?
Is it sufficient that the model weights (the associations and patterns developed through the training process) contain an abstracted representation of all the model’s training data?
The answer to these questions will depend on the court’s interpretation of UK legislation as much as the technology itself. As with all AI-related legal risks, it will be technology-led and the decision in this case may not be relevant to AI models trained using different techniques.
Is an AI model an infringing copy if its making constituted a copyright infringement or would have constituted a copyright infringement if it were made in the UK?
Another fundamental requirement is that Stability AI knew or had reason to believe that Stable Diffusion was an infringing copy. Is it sufficient that it is common knowledge that photographs are protected by copyright, and that copyright in them would be infringed by copying them, or by importation of infringing copies of them? Or did Stability AI hold a reasonable belief that Stable Diffusion was not an infringing copy because none of the model, source code, or output was a substantial reproduction of any copyright work from its training dataset?
Other AI developers need to pay close attention to the court’s ruling on these issues because it should clarify whether AI models trained in other countries will infringe under UK copyright law, and if so, how. The UK government in its recent Consultation on Copyright and AI specifically noted that it wanted to avoid UK-trained AI models from being disadvantaged compared to those trained elsewhere but operating within the UK.
Takeaways
International copyright disputes are never straightforward and those involving global AI operations are no exception. The Getty case has illustrated how important it is for a successful claim to establish an infringing act in the relevant jurisdiction, and this will depend not only on the specific techniques and datasets used to train the model, but also the differing exceptions to copyright infringement around the world.
The crux of the (UK) Getty case is now whether UK copyright law bites on an AI model that was trained elsewhere but made available to UK consumers. Even if most AI training currently occurs in the U.S. or China, the UK remains a valuable commercial market and developers wishing to sell into it need to take note.
This is at the same time, of course, as following the many ongoing proceedings in other markets (particularly the U.S., where much of the AI training takes place) as well as assessing the impact of the EU AI Act, which requires AI developers offering their products in the EU to have complied with EU copyright laws when training their models (even if trained outside the EU). It is a very complex and fast-moving global picture.
Footnotes
1. Thomson Reuters v. ROSS Intelligence No. 1:20-cv-613-SB (E.D. Pa. Feb. 11, 2025)—a summary judgment.
2. Andrea Bartz v. Anthropic C 24-05417 WHA
3. Richard Kadrey et al. v. Meta Platforms Inc. No. 23-cv-03417-VC
