A Mark Of Progress? Getty v Stability AI's Impact On IP Infringement In The Age Of AI

Whilst some commentators have described the first major AI case in the UK as a 'damp squib', the mammoth judgment of Dame Joanna Smith is rich with legal interpretations relevant to both the creative industry and those engaged with generative AI.

Getty has referred to the judgment as a "significant win" for IP owners, citing the finding of trade mark infringement and confirmation that responsibility for the use of such marks lies with the AI model provider and not the user.Stability AI has also claimed victory, stating that the judgment "ultimately resolves the copyright concerns that were the core issue", the company having succeeded on the one remaining copyright claim. Getty is now turning its sights overseas to the ongoing US litigation, where it says UK findings of fact will be used.

Although very fact specific, much can be learned from this long-awaited judgment, findings relating to the respective responsibility of platforms or users for generating synthetic images which infringe trade mark rights will have real significance for AI platform design, and build on other cases finding platforms liable for trade mark use (e.g.Swatch).

The court's analysis of copyright secondary infringement grapples with the difficulty of what is an 'infringing copy' in an age where digital reproduction fundamentally changes the format of a work. In some respects, the claims that were abandoned by Getty as the litigation progressed are as informative as those that were pursued to the conclusion of the trial and suggest commercial answers to some of the most burning infringement questions on model training and output liability. There are useful learnings both for businesses in their international enforcement of licensed copyright and for litigators in the challenges of obtaining utilisable and proportionate evidence.

The immediate impact of the judgment for AI in the UK is understated. The judgment suggests other avenues of dispute to be explored by a motivated claimant either willing to match Getty's substantial investment in confronting unlicensed AI developers or able to adopt a sufficiently discrete line of attack. Until that moment arises, attention will return to the government and their pending response to the Copyright and AI consultation (discussed here) for guidance on the fraught interplay between generative AI and rightsholders.

Background

In January 2023 (seemingly eons ago in the rapid legal landscape of AI) the stock photography service Getty Images brought intellectual property infringement proceedings in the UK against Stability AI, a generative AI company. Getty alleged infringement of their rights by Stability through the training, deployment and use of the AI model Stable Diffusion (the "Model"). The core argument made by Getty was the data sets used to train the Model included millions of images owned by, or exclusively licensed to, Getty.

The claim survived an early application for strike out and summary judgment in October 2023. The application focused on three key points:

• jurisdiction - were any acts by Stability AI complained of carried out in the UK;
• for the purposes of copyright, can an AI model be an article under s.27 of the Copyright Designs and Patents Act 1988 ("CDPA"); and
• was the appearance of a watermark in some of the synthetic images generated by the Model use in the course of trade by Stability of Getty's trade marks.

The recent judgment provides a thorough overview of how the Model works, digging into the technical operation of training and output. For an accessible summary of how generative AI works and image generators in particular, seeAI 101: How do AI tools work and why are lawsuits being raised?Given the process of training AI image generators is dependent on paired images and descriptions, Getty's vast bank of images with accompanying captions and metadata is inarguably a treasure trove for those engaged in AI training.

The dropped claims

The claims were significantly narrowed by Getty at trial, the most significant being the primary copyright infringement claims on the training of the Model and the synthetic outputs.

The training and development claim

Whilst of great interest to the IP community, the question of whether the training of the Model on copyright protected works was unfortunately not addressed in the judgment. While Stability accepted that at least some of the data used to train the Model included images from the Getty websites, Getty abandoned the training claim after evidence was tested during the trial on the issue of jurisdiction.

Copyright protection is territorial, even if the commercial exploitation of copyright works transcends national borders. Stability argued throughout, and Getty ultimately accepted at trial, that none of the training occurred in the UK. Potentially seen as the more straightforward of infringement claims for copyright, the abandonment of the argument by Getty demonstrates the centrality of jurisdiction to an arguable claim. The training and development claim is maintained in the US where it is accepted that the relevant datacentre used for training of the Model is located. Observers will be keen to see whether the US 'fair use' defences offer a safe harbour for AI training on copyright protected works.

Getty's secondary copyright infringement claim, discussed in detail below, was a novel attempt to overcome the territoriality of infringement. That claim which, if decided differently, could have resulted in all generative AI trained on non-licensed copyright protected material being an infringement when those models are downloaded in the UK.

Outputs claim

The outputs claim was another fascinating issue concerning the significant problem for the creative industry of competing with synthetic outputs that are highly similar to their original works. This claim was abandoned following Stability's adoption of the well-established principles of negative matching, preventing user inputs from generating outputs that appear on their face to be clear copies. A deliberate failure to implement such technical measures or filters to prevent an AI tool from outputting infringing works has led to other copyright owners commencing litigation against AI developers, including in the US dispute between Disney and AI image generation tool Midjourney.

While frustrating for this issue to have fallen away without a decision, it indicates how current methods for addressing user directed infringement remain a powerful defensive tool for service providers to limit litigation exposure.

Although of limited utility for copyright holders, the discussions on trade mark infringement and outputs does get into the detail of responsibility for copies of trade marks appearing in synthetic outputs.

Database rights infringement

The premise of the database rights claim was interwoven with the primary copyright infringement training and development and outputs claim. As a result of their abandonment by Getty, the database claim became inarguable.

Passing off

This claim was never abandoned by Getty. Getty argued the claim would stand or fall on the basis of the trade mark infringement claims. However, it only received a glancing coverage at trial and Dame Joanna Smith declined (with the parties agreeing) to give judgment on it. Although not explored in the judgment, the discussion surrounding trade mark infringement, discussed below, would arguably have applicability for misrepresentation. However, given the failure of the s10(3) Trade Marks Act 1994 ("TMA") claim, it appears unlikely that Getty would have met the low bar for establishing damage to sustain a passing off action.

Trade mark infringement

Users of the Model had been able to generate outputs that appeared to include watermarks similar or identical to trade marks own by Getty. Getty's position is that this was evidence of trade mark infringement under ss10(1), 10(2) and 10(3) TMA. The trade mark case appears, at the moment, the more significant aspect of the decision.

Threshold

The first question dealt with whether the volume of synthetic outputs that contained 'watermarks*' (a term used in the judgment to mean signs that looked to be watermarks appearing on Model generated synthetic outputs) met a threshold to consider infringement. As Stability did not advance ade minimisargument, the criteria applied was whether at least one user in the UK had generated a watermark*.

The conclusion varied depending on the version of the Model for which Stability was found to have made available. With earlier versions, the threshold was met (with some exceptions), whereas later versions had undergone sufficient adaptations, including filtering to specifically address the emergence of watermarks*, to mean that the judge could find "not one jot of evidence" that watermarks* were generated by real world users.

The threshold analysis highlighted the idiosyncrasies of evidence gathering. In seeking to generate an infringing output for the purpose of litigation, that evidence (whilst on its face supporting the argument advanced) may have little value when considering whether the same infringement would occur "in the wild" by "real world users". The conclusion by the judge on the lack of utility of the curated outputs meant that the application of the trade mark infringement was phenomenally narrow.

Average consumer

It was accepted, adopting the position from the Supreme Court in theIconixtrade mark case considering post-sale confusion, that the 'realistic and representative' way that watermarks* would be encountered would be when the image is generated by the user of the Model. Whilst Getty tried to argue that a post-sale average consumer should be considered, this was rejected as a watermarked* image is undesirable and was therefore unlikely to be shared with or seen by third-parties after initial generation. This conclusion was supported by Getty's own evidence of Reddit posters looking for workarounds to unintentionally generated watermarked* images.

The users of the Model could be broken into three average consumer categories depending on the technical skill used to access the Model. These categories were: (1) the user who downloads the codebase and model weights to run and adapt the Model offline; (2) the user who accesses the Model via an API or developer platform, and; (3) the user who accepts the terms and conditions to use the Model via a webpage. Irrespective of the method of accessing the Model, the judge concluded that the average user is likely to pay at least a moderate degree of attention to the image they have prompted, notwithstanding the intended use or the fact it may have been generated for free or at very little cost.

Context

The context in which an average consumer will encounter a sign is a relevant consideration. For Stability, a helpful context is one in which users understand that they are generating images on their own accord, that these images are 'synthetic' and that the Model cannot be expected to generate photorealistic images, what Stability termed 'out-of-scope use'.

A different contextual analysis was undertaken for each average consumer category. The category (1) consumer is the closest to Stability's ideal average consumer with the judge considering that whilst they are unlikely to have read the terms (as they were not required to accept them), they would understand that they are using "a cutting edge AI model to generate synthetic image using his own prompts and that the Model is not designed to produce out-of-scope images" and that any images generated are done so without the intervention of Stability. They will also, because of their technical skill, understand how the Model is trained and why watermarks* may appear. For consumer categories (2) and (3) they will have at least skimmed the terms they were required to accept (although none of those referred to out-of-scope use of the Model). Both will understand that the Model is provided by Stability and while consumer (2) will understand that the images are artificial, consumer (3) may only understand that images are AI generated in response to prompts.

S10(1)

S10(1) trade mark infringement requires there to be unauthorised use of a sign identical to a trade mark for identical goods or services in the course of trade. Here, the marks were the GETTY and ISTOCK trade marks and the signs compared with were some of the least distorted watermarks*.

To understand how an image isgenerated in response to a user prompt can be the AI model's use of a sign in the course of trade, it is helpful to consider what goods/services Stability is said to be providing. That is, the provision of synthetic image outputs as both a service (i.e. the using of the model) and a good (i.e. the synthetic image itself). With that in mind, the issue is the application of watermarks* to those goods, or via that service.

Commercial communication

Whether use of a sign is in the course of trade and by whom is helpfully encapsulated in the 'commercial communication' test and is more regularly applied in the context of platform use.

This argument boiled down to the degree of control exercised by Stability versus the user in the generation of watermarked* outputs. Stability had control over (a) the data input for the purposes of training (at least for many of the versions of the Model); (b) was able to exercise some control at the pre-processing prompt stage by flagging certain words or phrases that would need to result in an output that deviates from the standard process, i.e. prompts including celebrity names will result in modified outputs to address concerns with fake news and propaganda; (c) the post-processing output can also be altered to exclude harmful content that may have been generated in response to a prompt. This degree of control distinguished Stability fromGoogle France,CotyandL'Orealas it went beyond creating the technical conditions necessary for use of a sign and the users themselves had limited control in the output returned. As a consequence, the watermarked* synthetic outputs were Stability's commercial communication for some versions of the Model.

In relation to goods and services / in the course of trade

The judge found that these two infringement considerations were intricately linked by Stability's arguments and addressed them together. Stability put forward four prongs of defence. These were that an average consumer would: (i) consider that as users they were responsible for the outputs; (ii) consider the appearance of the watermark* to be a technical result rather than an indicator of origin; (iii)see the Stability branding and absence of standard co-branding from Getty and the "recognisably off" nature of the watermark* image as an indicator of no connection between the output and Getty; and (iv) if it was a genuine issue, users would have raised the concern with Getty, generating actual consumer evidence on which Getty could rely.

In summary, the judge found that average consumers would consider Stability to be offering the Model in the UK and affixing the watermarks* to the outputs via the operation of the Model. In effect, Stability were putting watermarked* images on the market as a commercial communication to the consumer. The user could not, and would not understand themselves to be, responsible for the watermarked* outputs because of several factors; in particular, the control Stability exerted over training and the fact that consumers would not intend to generate watermarked* images. Provided that the watermarks* were not overly distorted, an average consumer would consider the outputs to have originated either from Getty, or have been generated as a result of the use of licensed training data.

It should also be noted that written AI prompts which included the Getty word trade marks were expressly excluded from consideration. This decision implies that the parties and the court are less confident that intentional prompting could be the responsibility of the AI company. The sliding scale of responsibility is a point that could be very interesting in the context of outputs which are deliberately prompted to copy either trade marks or copyright protected works.

Identity of signs

Identity was found in respect of one pleaded watermarked* image for the ISTOCK trade mark.For the GETTY trade marks no identity was found in the pleaded images. Either they were not identical, or the watermarked* output was generated in a manner that would not occur "in the wild" and therefore had to be disregarded. Attempts were made by Gerry following circulation of the draft judgment to invite the judge to consider further images from the annex that were not relied on at trial, but this was rejected.

Identity of goods and services

The categorisation of Stability's goods and services were the provision of synthetic image outputs as a good and as a service. All access mechanisms described in the three categories of average consumer resulted in the provision of a synthetic image as a good and average consumer (2) and (3) would understand Stability to be offering the generation of synthetic images as a service. However, for average consumer category (1) the service was the provision of the code and model weights to be downloaded and worked on offline. Getty relied on broad goods and services relating to digital imagery, photography and image services. Some of these (although not photography) were found to be identical by construing synthetic image outputs (as a service and a good) as a new variant of the terms in the specification.

Conclusion on s10(1)

The judge made a finding of double identity infringement in respect of the ISTOCK trade mark based on only one pleaded example. Despite the narrow factual scenario, we consider this finding to be very consequential for the liability of LLMs that have the potential to generate outputs containing third-party trade marks.

S10(2)

S10(2) infringementrequires the signs and the marks to be identical or similar and the goods/services to either be identical or similar and as a consequence, there is a likelihood of confusion on the part of the public.

The decision on s10(2) overlapped to a large degree with the conclusions for s10(1). Infringement was found in the pleaded examples (other than the example of a watermarked* image that would not occur "in the wild"). The marks were identical or, where moderately distorted, similar to the watermarks*. As above, the goods and services were identical or similar. For the same reasons previously discussed relating to responsibility and the undesirable nature of watermarked* images, there would be a likelihood of confusion on the part of the average consumer as to origin of the goods, or commercial connection between the goods and Getty. A helpful example raised was that consumers would understand that a watermarked image usually requires a licence to use, and conclude, therefore that a watermarked* image would only have been generated with Getty's consent).

S10(3)

For s10(3) infringement a mark must have a reputation (this was not in dispute here) and the sign must be identical and similar, there is no requirement as to the nature of goods/services. However, use must be without due cause and have the effect of taking unfair advantage of, or causing detriment to, the distinctive character or repute of the mark.

Considerations on s10(3) focused on whether the watermarks* resulted in dilution or tarnishing of Getty's marks,or whether their use provided an unfair advantage to Stability. All arguments were rejected for two key reasons: lack of real life evidence of the appearance of watermarks*, and that the appearance of the watermarks* was undesirable to both Stability and the users of the Model. Getty's argument that watermarked* images may flood the market taking the place of Getty as a reliable photo provider was unsupported by any evidence of a change in economic behaviour. The argument that Stability was gaining some form of benefit by affixing the watermarks* to the images was inarguable given the evidence that watermarks* were intentionally being filtered out. The strongest of Getty's arguments was in relation to tarnishing, with the concern that use of the Models may produce a variety of damaging content including: not safe for work ("NSFW") watermarked* images; fake photorealistic watermarked* images; poor quality watermarked* images. Whilst Getty was able to produce NSFW watermarked* images of celebrities, there was no evidence that this had occurred "in the wild" and in the absence of any evidence on how likely it was for these potentially tarnishing images to appear, the judge was not prepared to make an inference of tarnishing. Likewise there was no evidence of watermarked* deep fakes occurring. In respect of the potential damage caused by distorted images the judge found that because the degree of distortion of both the synthetic image and the watermark* was incredibly varied, she was not in a position to make a conclusion.

Unlike most other trade mark cases, the appearance of an infringing sign in a synthetic AI output would likely be different in each circumstance and the resultant impact on the average consumer would be likewise varied. Taking a cautious approach to hypothetical damage to a mark seems eminently sensible in this circumstance.

Copyright

Secondary infringement

The only copyright claim to proceed to trial was on secondary infringement. Getty's case was that the model weights used in the training of the Model (the "Model Weights")were an "infringing copy" of a copyright protected work that was imported into the UK (s22 CDPA) or distributed otherwise than in the course of business in circumstances where Stability knew or had reason to believe that the Model Weights were an infringing copy of the work.

Model weights are the numerical encoding of patterns learned in the training data, i.e. how an image relates to text inputs.

The relevant definition for infringing copy is found at s27CDPA. To paraphrase: an article is an infringing copy if proposed or actually imported into the UK where its making would have constituted an infringement had that making occurred in the UK.

The discussion centred on statutory interpretation of whether the Model Weights can be an article and whether an infringing copy has to have contained a copy of the work at some point.

In consideration of the "always speaking" approach to legislative interpretation and on a review of the case law it was found that an article can be intangible. This conclusion is a pragmatic interpretation given that copies can readily occur entirely digitally, and potentially at all times in the cloud.

The question of an infringing copy was more complex. It was found that to be an infringing copy, it is not enough that there may have been infringement occurring during the 'making' of the article, despite the wording of the definition of an "infringing copy" in s27 CDPA, the article also has to contain (or be) a copy itself. The expert evidence stated that in the training of the Model Weights, works are reproduced, but they are then converted from a pixel space to a compressed latent space to ensure efficiency of training. The process of diffusion that creates the Model Weights occurs on the latent space. The Model Weights enable the Model to produce synthetic outputs. It was accepted by the court and the parties that the latent space representations of the works are not a reproduction of the work. Therefore, the Model Weights never contained a 'copy' and on the judge's interpretation cannot be an 'infringing copy'. That is irrespective of whether copyright infringement occurred at an earlier stage and was necessary for the making of the Model Weights and the Model.

Accepting perhaps that this was the most contentious area of the judgment and an area where there may be a future appeal, the judge went on to consider that if she was wrong on this interpretation and the Model Weights could be an infringing copy by virtue of their making, there would be secondary infringement for the versions of the Model Weights that require downloading for use in the UK.

Getty did not argue that the Model Weights contained a copy of the works, focusing on the theory that the making of the Model Weights would have constituted infringement if carried out in the UK. However, reproductions of works in a digital age undergo severe format changes for efficiency/practicality. Had this point not been accepted by Getty, it may have been necessary to consider at what point a 'copy' stops being a substantial 'copy' because of its medium and the processes applied to it. For example, is a 'copy' still a 'copy' if it is held in a binary format rather than when it is viewed, case law would seem to suggest it can be. In the case ofSony v Balla RAM memory chip that contained a transient copy of a copyright work could be an article and an infringing copy at the point at which it contained the work. Indeed, the experts agreed that the Model can generate "images that are nearly identical (a memorized image)" and even "images that are derived from a training image either in part or in whole (a derivative)", so the issue is not the similarity (or identity) of the synthetic image to the work, it is instead in the making where the actionable 'copy' is lost.

Entitlement

The section on entitlement was not the juiciest part of the judgment, but the analysis of a snapshot of Getty's various licence arrangements is a potent reminder to have a global perspective when drafting these agreements. Many of Getty's licences were prepared under New York law where it is possible for an exclusive licence to have multiple beneficiaries. Where a group company may wish to exploit licensed copyright throughout its various corporate arms, it's quite clear that a broader grant of licence would cut down on an internal administrative headache. However, when trying to rely on that licence to pursue a claim in other jurisdictions, the recipient of a New York law exclusive licence may find itself out of luck. In the UK, an exclusive licence, as defined in s92 CDPA, can only be granted to one party. Without that, the rights and remedies allowed to an exclusive licensee under s101 CDPA fall away. The consequence for Getty is that several of the exemplar licence agreements did not entitle Getty to take action on behalf of the copyright owner.

CONCLUSION

The shape of this case may appear very different than anticipated by the eagerly awaiting masses looking for guidance on AI and IP, with big unaddressed gaps on the liability under intellectual property law for AI training and generating outputs. However, its significance cannot be overlooked. In part, what was not argued or what was admitted is as illuminating as the substantive judicial analysis. It is also important to recognise that this decision represents only part of a broader cross-jurisdictional dispute between the parties, with Getty already indicating that it will be "taking forward findings of fact from the UK ruling in our US case".

For AI developers and deployers, this judgment offers some practical solutions to the risks posed by assimilating vast quantities of data: effective filtering, negative matching, training outside of the UK, offering image generation as a service in the UK with all relevant servers abroad, and avoiding the use of trade marks in outputs. For rightsholders it suggests arguments to be explored in future litigation: when is AI responsible for synthetic outputs (particularly for outputs containing trade marks, but perhaps this could extend to other rights), at what point does a digitally compressed/synthesised copy cease to be a copy for the purposes of copyright. Trade mark owners may also want to review the scope of their trade mark registrations, and consider whether they have protection for their signs in the relevant classes which would support analogous trade mark infringement actions against AI image generation companies which reproduce their signs in AI outputs.

As it stands, the decision has little bearing on how AI is currently operating in the UK. However, the potential for appeal to the finding on secondary copyright infringement could be tectonic. The judgment (and the issues that it does not address) will also increase pressure on the UK government as it grapples with potential policy options for resolving the conflict between AI developers and copyright owners. Under a compromise agreed to pass the Data (Use and Access) Act, the UK Government is required to produce an economic impact assessment relating to its Copyright and AI consultation, alongside a reporton the use of copyright works in the development of AI systems by Spring 2026. One key consultation question concerns the copyright status of AI models trained outside the UK, and the Getty Images v Stability AI decision provides a stark reminder of the importance of resolving that issue.

Read the full judgmenthere

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.