How AI Works, Matters

We have previously reported on how Artificial Intelligence (AI) as a rapidly evolving facet of modern technology has already had a significant impact on the way that we live our lives through the myriad of use cases that have emerged. We also posed the question of what happens when these systems are trained on copyright content owned by others and suggested that these complex issues will require further clarification by policy-makers and the courts in the coming years.

Well some of those answers have now been addressed by the UK High Court with the judgement in Getty Images v Stability AI [2025] EWHC 2863. In particular, if the AI was trained using copyright protected content, and if the AI subsequently stored or re-produced this content for the user to create outputs by manipulating with specific prompts, is the user or developer then infringing copyright?

In brief, Getty Images (one of the largest visual content libraries worldwide) alleged that Stability AI had trained its image-generation model (known as Stable Diffusion) on millions of Getty's copyrighted photographs. By the time the judgment was handed down, Getty had abandoned most of its claims. What remained was a narrow dispute over trade mark infringement.

How did the AI model actually work?

One of the key issues before the Court was how the AI image-generation model actually worked. The Court was unequivocal and held that the Stability AI system did not store or reproduce copyright works and determined that the model weights and probability distributions (the mathematical parameters that constitute the essence of an AI system) are not copies of training data. Justice Smith noted that the Stability AI system was "purely the product of the patterns and features which they have learnt over time during the training process."

Of course, that may not be true of other AI systems – and how AI works, matters!

In the agreed technical primer for the case, the experts discussed what is called 'memorisation', which is a major feature of the New York Times' claim against OpenAI as we have discussed previously in respect to direct copyright infringement of data used to train AI systems,¹ direct and indirect copyright infringement of AI generated outputs,²,³ and patent eligibility of AI generated content.⁴

The technical primer in the present case noted the following:

"The network's weights are optimised on the training data, but its goal is to perform well on previously unseen data. In the context of Stable Diffusion, unseen data means new random noise patterns and/or new text inputs. To work reasonably on such new data, the network must be able to 'generalise': to recognise and understand the general patterns and rules in the training data and be able to apply them in a different context.

Deep networks can both generalise and memorise at the same time. In such case, the network uses most of its weights to represent general patterns in the data, but uses some part of its weights to memorise individual patterns".

In relation to the question of whether or not models contain their training data or store their training data, the experts agreed that while Stability AI system does produce images that are distinct from the training examples, the model can also produce nearly identical images, i.e. a memorised image, or what is called in the New York Times and OpenAI case, a regurgitation, and that it can produce images that are derived from a training image, either in part or in whole (a derivative).

Primary Copyright Infringement

The Court accepted that while Stability AI had created datasets by scrapping the internet without consent of the owner of the copyright, the Stable Diffusion image-generation model did not store training data itself. Moreover, Getty Images also had difficulty establishing chain of title to many of the copyright images as the authors were employed by different entities with the Getty Group and could not establish standing as owner or exclusive licensee of the relevant works.

Territorial Limitation

Equally importantly in this litigation, notwithstanding the pleaded case, Getty Images acknowledged before closing that there was no evidence to support the claim that the training and development of the Stability AI system took place in the UK. It was also acknowledged that the type of prompts which Getty Images alleged to have been used to generate the infringement output from the model in evidence had been blocked by Stability. This meant that the relief to which Getty Images would have been entitled in respect of their allegations of primary infringement had no territorial basis and were abandoned. The primary reason for dropping these claims was that Getty Images was unable to prove that the training and development of Stability AI system took place in the UK. It is likely a similar result, on the facts, would be found in Australia.

Secondary Copyright Infringement

In its secondary copyright infringement arguments, Getty Images contended that Stability AI had imported into the UK an 'article', namely the Stability AI system. Getty Images did not actually say that Stable Diffusion itself was a copy or that it stores any of the copyright works. But they contended that the model was an infringing copy under the UK copyright legislation because the making of its model weights would have constituted infringement of the copyright works had it been carried out in the UK.

While the UK High Court found that the Stability AI system could be an 'article', it reasoned that an AI algorithm which does not store or reproduce copyright works cannot be an infringing copy – even if it was trained using reproduction of copyright. As the Stable Diffusion model did not retain a copy of the training data, the Court determined there could be no secondary infringement.

Take Aways

No doubt, this is just the tip of the iceberg in terms of the legal action to come on this particular issue. In Australia, the location of the jurisdiction in which the training was conducted and the datasets themselves, are likely to have a significant bearing on Australian court's jurisdiction over direct copyright infringement when using AI tools. Diligence to chain of title and corporate hygiene in demonstrating ownership of copyright and standing are given.

As this UK decision demonstrates, how AI works, matters!

Footnotes

¹ J.Doe et.al. v Github, Inc, Microsoft Corporation and OpenAI GP, LLC et al, 22 Civ. 6823 (N.D. Cal. Nov. 10, 2022) at https://githubcopilotlitigation.com/pdf/06823/1-0-github_complaint.pdf
² Naruto v. Slater, 888 F.3d 418 (9th Cir. 2018).
³ www.copyright.gov/docs/zarya-of-the-dawn.pdf
⁴ Commissioner of Patents v Thaler

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.