Getty Images v Stability AI: Key IP Issues in AI Litigation

The long-awaited Judgment in Getty Images v Stability AI has now been handed down by the English High Court (the full Judgment is available at https://lnkd.in/gHtVnHuY). What lawyers and rights holders hoped to be a landmark copyright action between a major rights holder and an AI developer was soon reduced to a trade mark dispute, with just a separate claim of secondary copyright infringement, which was unsuccessful. The Judgment has been viewed by many seeking judicial clarity on the most pressing copyright issues raised by the development and use of generative AI tools as little more than a damp squib.

Background to the case

Earlier this year, we previewed the English High Court trial between Getty Images, a leading visual content licensor, and Stability AI, the developer of the generative AI model Stable Diffusion. Getty alleged that Stability had used millions of its copyrighted images without consent to train its AI models, and that the output of these models reproduced substantial parts of Getty’s works and trade marks for its watermarks. The case was widely anticipated as a test of how UK intellectual property law applies to the training and operation of generative AI models, with the potential to set important precedents for both the creative and technology industries.

Key Issues in the Original Claim

Getty’s original claim was broad and ambitious, raising around 60 issues for the Court to consider. The main claims included:

• Primary Copyright Infringement: Getty alleged that Stability infringed copyright by (i) using Getty’s content to train Stable Diffusion; and (ii) by generating outputs that reproduced substantial parts of Getty’s works.
• Secondary Copyright Infringement: Getty alleged that Stability had imported and distributed an “infringing copy” of its works in the UK, contrary to sections 22 and 23 of the Copyright, Designs and Patents Act 1988 (CDPA).
• Database Right Infringement: Getty claimed that Stability had extracted and re-used a substantial part of its database in training the AI models.
• Trade Mark Infringement and Passing Off: Getty claimed that synthetic images generated by Stable Diffusion sometimes bore Getty’s or iStock’s watermarks, infringing its UK registered trade marks and constituting passing off.

The High Court Judgment

After a three-week trial, evidence from 19 witnesses and two experts, Mrs Justice Joanna Smith DBE handed down her gargantuan 205-page judgment on 4 November 2025. The outcome, however, was far more limited than many had expected:

Primary Copyright and Database Right Claims Abandoned

Getty abandoned its primary copyright infringement claims (in respect of both training and outputs) before closing submissions at trial. The training claim failed for jurisdictional reasons, as Getty could not show that the AI model was trained in the UK; while the output claim became redundant after Stability blocked the prompts that had allegedly been used to generate potentially infringing output. The parallel database right claims were also dropped.

Secondary Copyright Infringement Claim Dismissed

The Judge dismissed Getty’s claim for secondary copyright infringement.

Getty claimed secondary copyright infringement on the basis that Stable Diffusion is an “infringing copy” within the meaning of the CDPA. Getty alleged that Stable Diffusion is an “article” which Stability knew or had reason to believe is an infringing copy, which has been imported into the UK (through being downloaded by users in the UK) and the making of Stable Diffusion (had it been carried out in the UK) involved copyright infringement. Getty did not allege that the model itself was a copy or that it stored any copies of Getty’s protected copyright works.

The Judgment explains that, while AI models such as Stable Diffusion are trained on vast datasets, they do not store this training data in the model. Rather, Stable Diffusion learns the statistics of patterns in the training data and encodes those patterns numerically in model weights. This enables the model to recreate an image of something in response to a prompt by sampling from the probability distribution associated with certain concepts in the training data.

The Judgment rested on whether Stable Diffusion could be classed as an “infringing copy” i.e. an “article” (undefined in the statute) that has been imported into the UK. Getty contended that the term “article” should be interpreted broadly enough to encompass tangible and intangible objects, while Stability contended that an article can only be a tangible thing. Based on (i) the Judge’s reading of s.17(2) in light of Sony v Ball [2004] (in which copied data transiently stored in a physical article such as a silicone memory chip was capable of being an article); and (ii) the “always speaking” approach to statutory interpretation (in which a statute should be interpreted taking into account changes that have occurred since the statute was enacted), the Judge found that an article is intended to also cover electronic copies stored in an intangible medium, such as storage in “the cloud”.

However, in order to be an “infringing copy”, the article must still constitute a copy i.e. a reproduction of a copyright work. The Judge considered it was not possible for an article to be an infringing copy, if it has never consisted of, stored or contained a copy. In Sony v Ball, the memory chip only became, and remained, an infringing copy while it contained the copy data, i.e. after the reproduction had occurred and for as long as the chip held the copy. Given that Stable Diffusion’s model weights did not and do not store or reproduce any copyright works and have never done so, it could not amount to a ”copy” within the meaning of ss 22 and 23. The Judge rejected Getty’s arguments that Stable Diffusion became an infringing copy as soon as it was made (had that occurred in the UK) as the making would have involved the use of infringing copies.

Nevertheless, the Judge made a finding that downloads of Stable Diffusion onto devices in the UK did amount to importation within the meaning of the CDPA.

Trade Mark Infringement: Limited Success

Getty had limited success in its trade mark claim. The Court found Stability AI liable for limited acts of trade mark infringement, as UK users of certain, earlier versions of Stable Diffusion (v1.x and v2.x) could generate synthetic images bearing watermarks similar or identical to registered trade marks for “Getty Images and iStock”, when prompting these versions of the model.

Double Identity - s.10(1) Infringement

Did Stability use identical or similar signs to Getty’s trade marks?

The Court’s assessment of use of the Getty Images or iStock trade marks focused on whether Stability had control over the model and its generation of output affixing the Getty Images or iStock water marks.

Stability AI argued that users of the model had complete control over use of the model and the output generated in response to their prompts, including output bearing watermarks. However, the Court found that in being responsible for the model weights (which control the functionality of the model) for the Stable Diffusion models, writing the code downloaded by users accessing the models, training the model and making it available to users, Stability had done more than merely create the technical conditions necessary for the use of the Getty watermarks by users. By controlling on what data the model was trained (some of which contained images bearing a visible watermark) Stability was the only entity with control over the generation of watermarks on synthetic images produced by Stable Diffusion. The Judge found that while users understand that they generate images in response to their own prompts, the average consumer would not see themselves as solely responsible for outputs bearing watermarks, especially when such watermarks appear unintentionally (i.e. users do not seek to generate watermarks when prompting a model to produce an image as output). Rather, the appearance of watermarks is understood to result from the way the model was trained, for which users are not responsible.

The Judge rejected Stability’s argument that such use was outside the scope of trade mark law simply because the watermarks appeared after the consumer’s commercial decision to download, and access Stable Diffusion, concluding instead that the presence of the marks on output images could create the impression of a material link or connection between Getty Images and Stability in the mind of the average consumer. The Judge considered that users seeing the Getty watermarks on output would see the watermark as indicating that the image originates from Getty or is in some way connected to Getty, perhaps due to a licence agreement being in place between Getty and Stability for training of Getty’s works to develop the model. Stability contended that the watermarks generated on the synthetic images produced were distorted and displayed in a different way from how they are displayed on the Getty platform, such that a consumer would not associate them as originating from Getty or being the result of an agreement between Getty and Stability. However, the Judge considered that the average consumer paying a moderate degree of attention would not be alerted to this distinction.

Was the use of an identical sign?

The Judge found that, while there was evidence of signs identical to the iStock watermarks appearing in Stable Diffusion outputs, no evidence was produced of output with signs identical to the Getty watermark, other than as a result of experiments carried out by Getty’s legal counsel. No examples of “real world” use of the marks was identified. When combining this conclusion with a detailed assessment of the goods and services covered by Getty’s trade mark registrations, the Judge concluded that there was double identity infringement in respect of the iStock watermarks generated by users of the v.1.x model, but not in respect of other versions of the model and that there was no double identity infringement of the Getty watermarks in any of the models. Where no “real world” examples of a user generating an image bearing either type of watermark, double identity was not found with respect to those models.

There was no evidence of UK users generating the Getty watermarks using the later XL and v1.6 Models, so the trade mark infringement claims in respect of these models were dismissed.

Likelihood of confusion (s.10(2))

On the basis that the Judge found that Stability had “used” the signs and that consumers would assume a connection between the watermarks and Getty, and, perhaps, assume Stability had taken a licence to use Getty’s images, the Judge found infringement with respect to the iStock watermarks generated by earlier versions of the model. In addition, the Judge held that there was infringement with respect to Getty’s watermark on the basis that some of the blurred and distorted reproductions of the watermark in the model v2.1 output were similar to that watermark.

Reputation and unfair advantage (s.10(3))

In relying upon the recent ruling of Arnold LJ in Thatchers Cider Company v Aldi Stores Ltd [2025], Stability argued that Getty could not succeed on this ground as they had failed to plead that there had been a change in the economic behaviour of the average consumer of the goods or services for which the trade mark is registered.

Getty’s case was found to presuppose that users of Stable Diffusion would wish to use the model to access Getty’s licensed images without paying for it. However, no real-life evidence was produced to support this and the Judge was convinced by Stability’s point that a user seeking to avoid a licence fee in this way would still end up with an unwanted watermark. Accordingly, Getty failed to establish a change, or likely change, in consumers’ economic behaviour.

Getty was also found to have failed to produce any evidence of “real world examples” where its watermark had appeared on violent or pornographic images generated by Stable Diffusion in order to prove its claim of reputational damage. Additionally, the Judge was unconvinced by Getty’s claim that its reputation was damaged by the low quality of the images generated by Stable Diffusion with the Getty watermarks affixed to them, again, due the lack of “real world” evidence of the same.

The Judge was also unable to find any basis on which Stability sought to gain unfair advantage from watermarks appearing on synthetic images generated by its models.

Passing off

The Judge decided not to entertain Getty’s passing off claim due to its limited particularisation and as Getty accepted that its claim for passing off would ultimately stand or fall with the trade mark infringement claim

No Additional Damages

The Court found no basis for a claim of flagrancy or widespread infringement and would not award additional damages (for copyright infringement or more generally) .

Implications of the Judgment

The outcome of the case evidently highlights the complexity and sophistication of litigation involving AI technologies. Whilst the case has provided some clarity on the nature and training of generative AI models, it is disappointing that after three years of litigation and a three-week trial, we are no closer to definitive answers to the central questions of whether the use of protected content to train and develop an AI model constitutes copyright infringement (or whether any statutory defences might apply to such activity) or who is liable when AI models produce outputs which are similar to third party content.

It has opened one up one point (which may be appealed) that a trained model cannot itself be an infringing copy for the purpose of secondary copyright infringement, as it contains no copies of underlying works. This may have been decided on the facts of this case, but as a more general proposition, it may be challenged.

From a policy perspective, the judgment leaves the UK Government’s position unchanged. Earlier in the year, the Government had deferred proposals to revisit copyright defences in the context of AI, hoping that this case would provide clearer guidance. That clarity has not materialised, and it remains to be seen whether the Government will revisit the issue in the near future. Notably, the Australian Government recently dropped a proposal to introduce a fair dealing exception for text and data mining, citing concerns about weakening copyright protection.

While the outcome will be welcomed by the AI developer community, there is a growing recognition that commercial licensing deals may be the most pragmatic way to resolve the difficult issues raised by AI and copyright, providing certainty and reducing the risk of litigation. Otherwise, we may see a shift towards more trade mark-related claims in the AI context, as rightsholders seek alternative avenues for enforcement of their rights. However, this case highlights that these trade mark infringement claims are not without difficulty and their success will turn on the particular facts.

As far as the parties to this case are concerned, the story is not over yet. They still await the outcome of parallel infringement proceedings in the Northern District of California, and, whilst they await the outcome, UK commentators will be looking to see if there is an appeal to the Court of Appeal….