Can an AI patent stop a copycat in another jurisdiction?

A patent gives its proprietor a monopoly over a specific domain of technical subject matter, the scope of which is defined by the claims of the patent. The claims of a patent for an AI invention typically define the invention through a series of computing steps, or by specifying hardware arranged to perform them.

The protection conferred by a patent is territorial, such that in principle a patent is only infringed when a third party performs certain actions in the territory for which the patent was granted. For inventions relating to physical things like devices and chemical compositions, there is often no doubt as to where an infringing act has taken place. Conversely, where the invention is defined by a sequence of computing steps, it can be difficult to establish where those steps are being performed, or indeed to establish what it means to perform the method in a specific territory. This is particularly true in the case of AI inventions, where often virtually all of what makes the tool new and useful (such as its manner of training and its use in inference) is implemented remotely, with the only computing that is carried out on the end user’s device being the provision of a straightforward interface for transmitting inputs and outputs.

As an example, consider a UK patent that claims a sequence of steps including receiving a prompt, performing inference based on the prompt using a particular trained AI model, and transmitting the output of the inference. There is a suspected infringer who provides users with access to a similar AI model via the internet. There are users of the suspected infringer’s model situated in the UK, who provide inputs to the model and receive its outputs through an application run on their personal devices, while the model itself is facilitated by a remote data centre in another country.

In this example, even if it can be determined where the hardware that executes the model is located, the question arises as to whether all the steps of the method are performed in the territory of the patent since, while the user who benefits from the use of the model does so in this territory, the inference steps take place outside of it, within the remote data centre. The suspected infringer would wish to argue that, because the inference does not take place in the UK, they do not perform the patented method in the territory of the patent (and therefore do not infringe). On the other hand, we would expect the proprietor of the patent to feel that, because it makes no difference to the AI tool’s users where the server that executes the inference is physically situated, it would not be fair that their invention is protected only when this processing is performed by a machine located within one territory.

This issue was considered by the UK courts in the case of Menashe Business Mercantile Ltd. v William Hill Organization Ltd.  The defendant, William Hill, operated an online gaming platform, which the claimant, Menashe, argued infringed its UK patent. The platform was hosted on a server located outside the UK, in Antigua or Curaçao. The judge had to consider whether it was a defence to the infringement claim that the host server was not present in the UK. They found that the answer to this question was “no”, stating that “[a]ny other result would be monstrous – allowing a defendant to use supposed cross-border problems to avoid infringement of a system anywhere.”  This conclusion was upheld on appeal.

The Menashe v William Hill judgment suggests that, in the UK, an alleged infringer of a patent for an AI tool is unlikely to be able to escape an infringement claim merely by showing that the tool is hosted by a machine located in another territory. However, while the issue of cross-border infringement in relation to computer-implemented inventions has been addressed by courts some other jurisdictions, there is no harmonised international position on this issue. The basic position in patent law internationally is that patents are territorial in nature, and it therefore cannot be assumed that the same conclusion will always be reached in other jurisdictions.

An exacerbating factor in demonstrating infringement of a patent relating to an AI model is that it can in practice be difficult to establish where an AI model hosted by remote computing infrastructure is in fact being executed. This was demonstrated in recent litigation between Getty Images and Stability AI, where Getty Images abandoned a copyright claim relating to Stability AI’s alleged use of its images in training an AI model in light of being unable to establish that this training, which was performed on Amazon Web Services infrastructure, took place in the UK. 

The absence of a settled international position on infringement of computer-implemented inventions across borders presents a challenge when designing a strategy for robustly protecting an AI tool. The rapid rise of the AI industry suggests that this is an area of law that may develop in future, and this indeed may motivate legislators to consider changes to the existing treaties and statutes. In the meantime, AI companies should work with their IP advisors to develop strategies that account for this uncertainty, considering where infringement might take place and how it would be proved when enforcing their rights.

For further discussion or advice on patent protection in AI or digital health technologies, please contact us at gje@gje.com