1st UK Judgment on generative AI and IP – Getty Images v Stability AI - Fox Williams

The High Court has provided the first UK judgment addressing alleged intellectual property (IP) infringements arising from the use of generative artificial intelligence (AI). Getty Images (a global image and stock photo licensing agency) and related companies brought a claim in the High Court against Stability AI (the developer of an AI image-generation model called Stable Diffusion). The 200-page decision is fact-specific, but provides valuable guidance on how existing copyright and trade mark law applies in the AI context.

Getty Images alleged that Stability AI had used its images to train its AI model without permission. It sought to establish that both the training of the model and the images generated infringed its IP rights:

Primary copyright and/or database rights infringement – Getty Images claimed that its images were unlawfully used to train and develop Stable Diffusion, and that some of the resulting generated image outputs reproduced substantial parts of its protected works. These claims were withdrawn due to the absence of evidence that the relevant training/development happened in the UK, and because Stability had modified its model to prevent the specific prompts that produced the allegedly infringing images. Secondary copyright infringement – Getty Images argued that Stability AI had imported into the UK, possessed in the course of business, offered for sale, or distributed, an “article” (Stable Diffusion) which it knew or had reason to believe was an “infringing copy”. The distinction from primary infringement was that secondary infringement concerned dealing with or making available allegedly infringing works (e.g. by providing access to an AI model that can generate the images for the user), rather than Stability AI itself directly carrying out an act of copying. Trade mark infringement and passing off – Getty Images claimed that AI-generated image outputs unlawfully reproduced its watermarks for “Getty Images” and “iStock”, which were protected via trade marks. The parties decided not to provide additional arguments to the Court for considering passing off, so no finding was made on that claim.

Secondary copyright infringement

As explained above, Getty Images argued that Stability AI had imported into the UK, possessed in the course of business, offered for sale, or distributed, an “article” (Stable Diffusion) which it knew or had reason to believe was an “infringing copy”. A key and new issue was whether a generative AI model could fall within the meaning of an “article” and could be an “infringing copy”.

The Court accepted that, in principle, an intangible item (such as a trained AI model) could constitute an “article” for the purposes of copyright law. However, the article still needed to be an “infringing copy” of the copyright work. The Stable Diffusion model itself did not copy and had never stored copies of the individual training images used during its development. It instead used and learnt from statistical representations (known as ‘weights’) derived from such images. As a result, on the facts, there was found to be no copying capable of giving rise to secondary copyright infringement.

Getty Images relied on GETTY IMAGES trade marks (two word marks and one logo mark) and two ISTOCK word trade marks for its watermarks and argued that the Stable Diffusion model used identical or confusingly similar signs (by reproducing its watermarks in the generated images).

The Court referred to specific examples (see below) of watermarks generated by Stable Diffusion and held that Getty Images succeeded (on the specific facts) in relation to older versions only (Stable Diffusion v1.x, v.1.2 and v2.1). However, the judge commented that the “findings are both historic and extremely limited in scope” and that it was “impossible to know how many (or even on what scale) watermarks have been generated in real life that would fall into a similar category”.

This finding did not extend to later versions (such as Stable Diffusion XL), as there was no evidence that users in the UK had generated images reproducing the Getty Images or iStock watermarks.

In summary, Getty Images secured a narrow win limited to older versions of Stability AI’s model. The decision shows that existing IP law can address AI disputes, but this depends on the evidence (which may be difficult to obtain for AI models where third party users had entered prompts in the past to generate an image). There are still many questions left to be answered under UK law, especially regarding primary copyright infringement where AI models are trained within the jurisdiction and potentially in relation to other types of generative AI models. We provide below some key takeaway points:

The judgment included that “Both sides emphasise the significance of this case to the different industries they represent: the creative industry on one side and the AI industry and innovators on the other. Where the balance should be struck between the interests of these opposing factions is of very real societal importance”. As mentioned in our previous AI article, the UK government held a public consultation on AI and copyright and we are waiting to see what (if any) changes are made in this area to overcome the difficulty of balancing protecting the rights of creatives with continued innovation in the AI sector.

An AI model can, in principle, give rise to secondary copyright infringement (e.g. importing into the UK, possessing in the course of business, offering for sale or distributing, an infringing article into the UK). However, this depends on whether the AI is using an infringing copy of the protected copyright work. In this case, Stable Diffusion did not store copies of training images, so copying was not found.

Developers potentially can be liable for trade mark infringement in outputs from generated AI models. On these facts, Stability AI was found to have infringed trade marks of Getty Images in older versions only of Stable Diffusion, as there were examples of images using the protected watermarks.

Where the training or development of an AI model takes place can impact where a claim can be brought for infringement. In this case, primary copyright infringement and database rights infringement.

This case shows the value of using watermarks, as if the generative AI reproduces them then this could be evidence of infringement.