Shaping AI’s Legal Future: The Implications of the Getty Ruling for Training Data Practices

The High Court’s decision in Getty Images v Stability AI marks the UK’s first major ruling on how intellectual property law applies to generative AI. While Getty secured a narrow trade mark win, the Court’s rejection of broader copyright and database claims, especially in its analysis of whether model weights can constitute “infringing copies”, offers an early blueprint for how UK law may treat AI training practices, leaving significant questions for future litigation and appeal.

Getty Images vs. Stability AI is the first fully‑argued UK decision between a major global rights owner and an AI developer – in this case, one behind a widely‑used text‑to‑image model, Stable Diffusion. The judgment marks an important but measured step in the legal framing of generative AI: the English High Court largely rejected Getty’s copyright and database claims, found a narrow instance of trade mark infringement in early Stable Diffusion outputs, and provided a detailed analysis of secondary copyright infringement that many will see as a roadmap for the legitimacy of AI training practices going forward. While not the dramatic “AI on trial” moment some anticipated, it clarifies certain aspects of how IP law is applied in the context of AI training, though with many more questions yet to be answered.

The Data Behind the Dispute: How Getty Images Entered the Training Pipeline

Getty, a leading stock‑image and licensing platform, controls a vast curated database of photographs, video and illustrations, accompanied by rich metadata and distinctive GETTY IMAGES and ISTOCK trade marks and watermarked previews. Stability AI, founded in 2019, supports development of open‑source generative models, including Stable Diffusion, a leading text-to-image generative AI tool.

Getty argued that Stability AI had used millions of copyrighted images (including some prominently bearing Getty or iStock trade marks) to train its models without authorisation.

Getty’s core business revolves around licensing a vast library of images, curated over decades, and associated with its well-known marks. Stability AI, by contrast, operates at the cutting edge of open-source generative AI, making its model widely available both through paid platforms like Dream Studio and as downloadable software for use and adaptation by others. Getty’s primary claims centred on the scraping and copying of its images at scale, not just for training but also in situations where synthetic watermarks and branding could be seen in AI-generated outputs, though as explained below, these were cut back by the time the case reached trial.

Procedural Background

At the outset, Getty pursued a wide range of different claims against Stability, including:

•  Primary copyright infringement - asserting direct, unauthorised reproduction of Getty’s works, especially in the model’s training phase.
•  Database right infringement - alleging extraction from and re-utilisation of Getty's uniquely curated image databases.
•  Trade mark infringement - arguing unauthorised use of registered marks, e.g., as watermarks.
•  Passing off - claiming that Stability’s system was misrepresenting its output as legitimate Getty content.
•  Secondary copyright infringement - related to importing or distributing allegedly infringing model weights within the UK based on an argument that the model itself should be treated as an “article”.

As the case advanced, Getty narrowed its claims. It withdrew the primary copyright infringement claim, accepting that core model development and data collection occurred mainly outside the UK, and also withdrew the bulk of its database rights claim. By June 2025, the court was left to consider: trade mark infringement, passing off and secondary copyright infringement (specifically whether specifically whether providing or distributing model weights trained on infringing works outside the UK could give rise to liability within the UK if those model weights are used or sold domestically).

Stability AI defended itself by asserting that any relevant copying took place entirely outside the jurisdiction of the UK courts, and that it merely provided a tool for user-initiated image creation. The company also suggested that only users, not Stability AI, could be considered infringers at the output stage; that images produced by Stable Diffusion were not sufficiently “substantial” reproductions to constitute infringement; and, finally, that any appearance of Getty marks was produced only by wilful manipulation, not through normal use of its services.

The Decision

Trade Mark

The High Court found limited infringement under sections 10(1) and 10(2) in relation to specific early versions of Stable Diffusion, on the basis that some outputs reproduced identical or highly similar GETTY/ISTOCK word marks (or figurative marks) in a way likely to be linked to the registered marks in the minds of average consumers. However, the section 10(3) claim was rejected, with the Court concluding Getty had not shown sufficient evidence of a change in the economic behaviour of the average consumer, or serious risk thereof, nor proven unfair advantage, dilution or tarnishment flowing from the relatively rare and degraded watermark outputs.

The trade mark success was therefore extremely limited - a modest win for Getty tied to specific, largely superseded instances of model behaviour, not an ongoing bar on offering Stable Diffusion in the UK going forward by any means.

Passing Off

Getty’s passing off claim was rejected. The Court held that the classic elements of goodwill, misrepresentation and damage were not satisfied: in particular, the evidence did not show consumers being misled into believing that AI‑generated images were genuine Getty content, or that Getty’s business had suffered the sort of damage passing off is designed to prevent.

Secondary Copyright Infringement

The secondary copyright claim was the most novel aspect of the case and is likely to attract the greatest long‑term interest. Getty did not contend that Stable Diffusion stored copies of its images; instead, it argued that the model, specifically its learned weights, constituted an “article” that was an “infringing copy” under section 27(3) Copyright Designs and Patents Act 1988 (CDPA) because the making of those weights, if done in the UK, would have infringed copyright in millions of Getty works. The Court undertook a detailed construction of “article” and “infringing copy”, examining how far these concepts can extend to machine‑learning models whose parameters encode statistical relationships derived from copyright works but do not reproduce those works themselves.

It accepted expert evidence that Stable Diffusion does not store training images, that inference runs without reference to training data, and that the model weights embody learned correlations rather than compressed copies of particular works. On that basis, the judge concluded that a set of model weights for a diffusion model of this kind is not an “article” which is itself a copy of the underlying works, within the meaning of the CDPA. The secondary copyright claim therefore failed in its entirety.

“The Court offered clarity on some points and caution on others, a reminder that the legal framework for generative AI is still taking shape.”

- Chris Sleep, Head of Litigation & Dispute Management