Getty Images v Stability AI: a landmark trial for generative AI in UK?

The Getty Images v Stability AI trial was heard in the High Court before the summer recess and it did not disappoint. From abuse of process to appeals, the trial threw plenty of curveballs.

Fundamentally, the case arose from one of the defining commercial legal issues of the moment – the contested intersection of intellectual property (IP) rights and artificial intelligence (AI). When it arrives, the High Court's judgment will make for a fascinating read, though it will not answer the question many were hoping for – whether or not the web-scraping of online content, and subsequent use of that content to train an AI model in the UK, is a primary infringement of copyright or database rights in the UK. This is because no evidence came to light in the trial (or in disclosure) that any of the development and training of the Stable Diffusion model occurred in the UK.

What are some of the key issues arising from the dispute?

AI training

Getty Images, a leading stock image provider, accused Stability AI of using its copyright-protected images without permission to train its AI model, Stable Diffusion. This model generates images in response to text or image prompts. It was trained on over 12 million images scraped from websites and other online content.

Getty claimed that Stable Diffusion's training dataset included images copied from Getty's websites without its consent, ultimately leading to the Stable Diffusion systems generating images which reproduced Getty's images and trade marks.

Initially, Getty pursued claims for primary and secondary copyright infringement, trade mark infringement, passing off, and database right infringement.

Initial copyright claims

The main Getty claims were of primary infringement of copyright on the basis of alleged unauthorised use of its images in the training and development of Stable Diffusion and in the output images that the system generated in response to user prompts.

However, Stability and its witnesses said that all training took place outside the UK, including on cloud servers located outside the UK, and Getty struggled to point to anything concrete to show otherwise.

Eventually, after evaluating Stable Diffusion's evidence, Getty dropped the primary copyright and database right infringement claims, saying that it did not have sufficient evidence on the details of the development and training process, and so could not prove that infringing acts during the training process occurred in the UK.

Getty likewise had difficulties in showing that the output images infringed, in the face of defences from Stability that there was no copying of substantial parts of the underlying images and/or that Getty did not own them, and/or that the images were elicited only when Getty using highly contrived prompting.

Memorisation and derivation

Getty argued that Stable Diffusion effectively memorised its watermarks during repeated training on images bearing these marks, allowing the model to generate images with Getty's watermarks even without prompting that explicitly mentioned Getty.

Stability countered that its AI models' weights do not store relevant information from individual training images. Its expert witnesses agreed that the model weights do not directly store such values, supporting its position that the models do not memorise training data in a way that constitutes copying. Stability explained that, during training,  the models learn patterns and features based on multiple images rather than memorising any specific images.

Ownership and licensing

Getty faced challenges over copyright ownership and exclusive licensing of works created by its photographers. Key disputes included whether the photographers were employed by the correct Getty entities and whether their contracts were effective to assign copyright.

Stability pointed to inconsistencies and gaps in the chain of title and questions over the extent of exclusivity. The case amply demonstrated the virtues of having clearly drafted contracts and good record keeping, with Getty being obliged to defend its position by producing human resources records, service contracts and witness testimonies.

Narrowed claims

Getty claimed Stability committed secondary copyright infringement under sections 22 and 23 of the Copyright, Designs and Patents Act 1988 (CDPA). Secondary infringement involves various acts of dealing with infringing copies of a copyright work without the copyright owner's licence, including importing them or possessing them in the course of business. In each case infringement requires that the infringer "knows or has reason to believe" that the article in question is an infringing copy.

It is worth revisiting the actual wording of section 27 of the CDPA here, which defines an "infringing copy":

"(1) An article is an infringing copy if its making constituted an infringement of the copyright in the work in question.

(2) An article is also an infringing copy if –

(a) it has been or is proposed to be imported into the United Kingdom, and

(b) its making in the United Kingdom would have constituted an infringement of the copyright in the work in question, or a breach of an exclusive licence agreement relating to that work."

Getty argued that the Stable Diffusion system effectively contains copies of Getty's images, and that making it available for download in the UK constituted importation of infringing copies, and hence infringed its copyright.

Stability contended that the Stable Diffusion systems do not store or reproduce visual information from training images, and so cannot be infringing copies. Stability also denied knowledge of infringement, asserting its belief that the models do not reproduce Getty works is reasonable based on the technical evidence.

Getty pointed out that the definition of "infringing copy" does not require that the article actually reproduces the copyright work; the mere fact that its creation would have involved infringing copyright was enough, and that this was the case even if the Stable Diffusion models were trained outside the UK.

Trade mark infringement and passing off

Getty claimed significant goodwill in its trade marks which it says are associated with authentic, high-quality photography. Getty argued that the inclusion of its marks on the images generated by Stable Diffusion would be likely to mislead consumers into believing these images are endorsed by Getty and that there is a "commercial connection" between Stability and Getty.

Getty presented employee testimonies and evidence of Stable Diffusion generating images which showed Getty's watermarks, highlighting the potential for this to cause harm to the Getty brand (for example, if watermarked images showed illegal or adult-themed content).

In order to succeed in an infringement claim, Getty would need to show that the appearance of watermarks of its brand on generated output images amounts to use of Getty's marks "in relation to" the outputs, and that such use was "in the course of trade and likely to lead to consumer confusion as to trade origin".

Stability argued that watermarks are unintended artefacts of the model training process and that consumers understand that they are mere technical artefacts, rather than commercial communications. Stability pointed to the absence of consumer complaints or confusion as evidence that the watermarks do not mislead consumers as to the images' trade origins.

Osborne Clarke comment

Getty framed this case as a "bet the farm" litigation, not just for itself but for creative industries as a whole. The argument is that if the creators of the training data are not compensated for the use of their content to train AI models, then this poses an existential threat to the creative industries, due to the competing outputs which can be produced by the models.

The jurisdictional issue about whether the development and training of the Stable Diffusion model occurred in the UK has always had the potential to derail Getty – it is established law that IP rights are territorial and for any act to be infringing, it must take place in the relevant territory. Getty drew inferences from the fact that Stability AI developers lived and worked in the UK, maintaining that therefore they "must" have downloaded Getty content onto servers in the UK, but these arguments were not enough to advance the case.

Getty's last remaining hope to establish a precedent for the creative sector hangs on an unorthodox interpretation of secondary copyright infringement law – that the importation of an "article" into the UK could include not just a physical thing but an intangible thing, namely the pre-trained Stable Diffusion model.

If the court sides with Stability on this issue, and rules that AI systems are not infringing copies, then UK rightsholders will struggle to enforce their copyright – whether in text or in images – in cases where model training takes place overseas.

However, if the court finds for Getty, offering rightsholders a route to enforcement in front of UK courts, it might affect where AI developers choose to set up shop. While these possible economic ramifications are not directly relevant to the judge's deliberations, they will feed into the UK government's response to its consultation on copyright and AI, which raised exactly this issue.

Whatever judgment is handed down, we can expect an appeal to the Court of Appeal, so this is unlikely to be the end of the matter.