AI training on trial: What Getty Images v Stability AI means for AI developers and copyright holders

Earlier this month saw the eagerly awaited judgement handed down in Getty Images and others v Stability AI [2025] EWHC 2863 (Ch).

For many it was hoped that the case would bring much needed clarity to the intersection of copyright law and generative AI, in particular the use of copyright materials in the training in generative AI.  However as we discuss below, whilst some clarity was provided, it seems key questions remain unanswered.

The case raises practical questions about lawful data sourcing, the scope of copyright protection in training datasets, and the sufficiency of safeguards when deploying AI models in commercial settings.

Background and claims

The main dispute revolved around the allegation that Stability AI used vast numbers of Getty Images’ photographs, thumbnails and associated metadata to train its Stable Diffusion image‑generation model without a licence.

Getty were of the view this involved large‑scale copying of protected works from its platform and from datasets compiled using web‑scraped content, and that the copies made during dataset creation, preprocessing and model training infringed copyright and related rights.

Getty also objected that the use of its curated collections amounted to unauthorised extraction from a protected database, and that the removal or alteration of watermarks and other copyright‑management information, as well as the appearance of distorted “Getty Images” marks in some model outputs, engaged claims relating to rights management and trade mark/passing off.

Getty’s position is that its catalogue is a commercially licensed library and that high‑quality, rights‑cleared images are a core input it monetises; unlicensed use to train a competing generative model, in its view, harms that licensing market and free‑rides on the investment made in creating and curating the collection.

Stability AI argued that training an AI model does not substitute for the images themselves, that any copying was incidental to technological processes, that data was sourced from the open web or third‑party datasets rather than from bypassing Getty’s systems, and that outputs are generated rather than reproduced.

The case therefore raised foundational questions about whether and when copying for AI training requires permission, the scope of any applicable exceptions, and the legal significance of watermarks and other identifiers appearing in outputs.

Findings in relation to the main claims

Training/development claim (primary copyright infringement)

In relation to Getty’s primary copyright claim, that Stability AI had copied their protected works in training and developing Stable Diffusion, this was dropped shortly after the trial begun, as they could not show that such infringing acts occurred in the UK.

Importation of an “infringing copy” (secondary copyright infringement)

In relation to Getty’s secondary copyright infringement claim, that the Stable Diffusion model, (its model wights etc) constituted an “infringing copy” of Getty’s copyrights works under the Copyright, Designs and Patents Acts (CDPA) 1988.

The Court held that it does not.  As set out at paragraph 600 of the judgment “the model itself does not store any of those Copyright Works; the model weights are not themselves an infringing copy and they do not store an infringing copy. They are purely the product of the patterns and features which they have learnt over time during the training process.”  In other words they are just statistical parameters derived from training.

Accordingly since the model never “contained” a copy of the copyrighted works it cannot be in and of itself an “infringing copy” under the CDPA.  The secondary copyright infringement claim must therefore fail.

Trade mark infringement

Getty had more success on this albeit limited.  The Court found that in some earlier versions of the model, and under certain prompts, outputs were generated that included Getty’s or iStock watermarks (which are registered trade marks).

However whilst Getty Images succeeded (in part) in their trade mark infringement claim the Judge found that such infringement to be both historic and extremely limited in scope.  The claim under s10(3) Trade Mark Act 1994 (relating to detriment to distinctive character or reputation. unfair advantage) was dismissed.

Passing off

The Judge declined to address Getty Images’ allegation of passing off in light of its limited success on its trade mark infringement claims.  As the Judge states at paragraph 542 of the judgement “I do not presently see that its claim of passing off adds anything to my findings on those [trade mark infringement] claims.”

Comment

The dropping of the primary infringement claim was a huge blow and means that the fundamental question of whether training and developing generative AI constitutes copyright infringement under English copyright law remains unanswered.

Whilst the secondary infringement claim failed, it is important to note this was highly technical and turned on the facts in question.  In particular it relates to the Stable Diffusion model which is particular type of generative AI.  The same conclusion may therefore not apply to other different types of generative AI and so developers should be alive to this fact.

The risk around misuse of watermarks / trade marks remains a real risk to AI developers who will need to ensure they have in place relevant checks and balances in place to ensure this risk is minimized.

From the creators / AI copyright holder’s perspective, the lack of success of the copyright infringement arguments in this case should serve as a warning that reliance on IP rights (particularly copyright) alone is not enough to ensure their copyright works are used in the way they would like. The need to monitor and police where their copyright works are being used and made available is key and to ensure suitable robust contractual licensing terms in place with licensees to help control and restrict use of their works is now more important than ever before.

How we can help

If you would like to speak to one of our AI experts or make an enquiry about the AI toolkit, please use the contact form available at AI governance toolkit.