Getty Images v Stability AI: The Great British Scrape-Off

In a landmark legal battle, Getty Images challenged Stability AI in the High Court, accusing the creators of scraping millions of copyrighted images to train their generative AI.

Introduction & Background

Readers will recall that in January 2023, Getty Images brought a High Court claim against Stability AI, alleging that the developer of the generative AI model Stable Diffusion, had infringed its intellectual property rights. Getty Images’ original claims were wide-ranging, including primary and secondary copyright infringement, database right and trade mark infringement as well as passing off. The heart of Getty Images’ case was the allegation that Stability AI had used around 12 million of Getty Images’ copyrighted images, scraped from its websites without consent, to train its AI models.

Claims narrowed

As the proceedings progressed, Getty Images was forced to abandon several key aspects of its claim. Notably, it conceded that there was no evidence the training and development of Stable Diffusion took place within the UK; Stability AI had suggested that this took place on cloud servers abroad. Getty Images similarly dropped the claim regarding infringing outputs generated in the UK after such prompts were blocked by Stability AI thereby depriving Getty Images of a remedy. It also withdrew the database rights infringement claim, which was inherently linked to the abandoned copyright claims. This left the court to consider only the narrowed issues of secondary copyright infringement, trade mark infringement and passing off.

The Court’s finding

On the remaining issues, the court found that Getty Images’ claim succeeded only in part. The court dismissed Getty Images’ claims regarding the secondary copyright infringement claim. The judge ruled that, although an “article” for copyright infringement purposes could be intangible, an AI model like Stable Diffusion, which fundamentally does not store or reproduce any copyright works, cannot be considered an “infringing copy” under UK law.

Trade mark tussles

The judge held, however, that Stability AI had infringed Getty’s ‘ISTOCK’ trade marks but only in respect of certain versions of Stable Diffusion (v1.x) and only where accessed via Stability’s own platforms (DreamStudio and/or the Developer Platform). The court also found limited infringement of the Getty Images marks in relation to outputs generated by v2.x. The court dismissed the claims in relation to detriment to the distinctive character or reputation of its marks and passing off.

The copyright question remains

Although the case had been anticipated as a significant development in the debate around the methods of training AI models, the discontinuation of the primary copyright and database rights claims has diminished its impact. The question as to whether such activities constitute primary copyright infringement is still, therefore, unanswered. It follows that AI companies should not rejoice, nor should the creative industries be too despondent.

Looking ahead

We wait to see if Getty Images will appeal the decision. In the meantime, focus will likely turn to the government’s consultation on copyright and AI. Here, the government’s preferred option had been to provide for a text and data mining exception, potentially with an opt-out and transparency requirements. The consultation engendered a significant response, with the creative industries expressing concern with the lack of fair remuneration. Following political pressure in relation to the Data (Use and Access) Bill, the government was subsequently forced to make concessions around a timeline for progressing the development of its policy. As a result, the government has established working groups on four themes: control and technical standards, information and transparency, licensing and wider support for creatives. These working groups are to meet over the course of the next couple of months.

Conclusion

The case highlights the practical and evidential difficulties faced by rights holders in bringing such claims, particularly in relation to jurisdiction and proving where and how AI models are trained, and in linking specific outputs to specific infringements.