Getty Images v Stability AI: where are we after the trial - trade marks?

What did the parties argue, what are the key issues and to what extent will the findings be relevant to other trade mark infringement claims involving AI?

What has happened?

• Getty led at trial with its trade mark infringement claims, reflecting the importance these claims took on as the case developed and its copyright claims stumbled (more here). 

• Getty relied solely on infringement at the output stage. It did not argue that the training and development by Stability of its models infringed Getty's trade marks (as it can be more difficult to show a relevant type of use or damage at that stage). 

• One of the key issues in dispute is whether the average consumer views the watermarks as indicating trade origin of the outputs or (alternatively) as unwanted artefacts of the AI training process. This is relevant to a number of issues in dispute, in particular, whether or not there is use of Getty's marks "in relation to" the outputs (which requires trade origin use) and whether or not there is a likelihood of confusion (which requires confusion as to trade origin).

• If the judge finds that the average consumer views the watermarks as mere artefacts of the AI training process, all of Getty's trade mark infringement claims should fail (since there will be no use of Getty's marks "in relation to" the outputs, as required for infringement). Conversely, if she finds that they will be viewed as indicators of origin, there is a reasonably good chance that Getty's claim of a likelihood of confusion will also succeed (since the issues are similar). Getty's reputation-based claims are more finely balanced.

• If Getty wins on infringement, it could be a pyrrhic victory if Stability is able to filter out use of the Getty watermarks, as it is trying to do. The same is true if Getty wins only in relation to some – but not all - versions of Stability's model, which is quite possible on the facts. 

• If Getty does not win, it does not necessarily mean that all trade mark infringement claims involving AI will fail. The unintended use of a third party's trade mark as a watermark might be different to, for example, the use of a third party trade mark on a virtual product within an output. 

Want to know more?

The trade mark claims

Getty alleges that Stability's model generates synthetic outputs containing the words "Getty Images" or "iStock" as watermarks, both of which are registered by Getty as word only and logo trade marks in the UK. Getty contends that this constitutes direct infringement under section 10(1), a likelihood of confusion under section 10(2) and unfair advantage/detriment under section 10(3) of the Trade Marks Act 1994 (TMA). It also alleges passing off, but acknowledges that this claim stands or falls with the trade mark infringement claim. Examples of watermarked images alleged to infringe are below.

Getty says that Stability is primarily liable for these infringements. There is no allegation that Stability is liable as an accessory (ie no allegation that there is a common design to infringe between Stability and its users and/or that Stability has procured infringements by its users).

Arguments common to all claims

All acts of infringement, irrespective of the section under which they arise, require that a person uses in the course of trade a sign in relation to goods or services without the consent of the proprietor. All claims also rely on the perceptions – and level of attention – of the average consumer. We consider each of these issues first, then the specific allegations of infringement relied upon by Getty.

Is there use?

The TMA sets out a non-exhaustive list of examples of use that can constitute infringement, including affixing the mark to goods or packaging and offering or exposing goods for sale under the mark. Getty argues that this is exactly what Stability is doing - affixing the allegedly infringing marks to its outputs and offering those outputs to users – so there is no need to even go outside of the wording of the TMA when considering whether there is use. Stability says that it is not using the Getty marks in any of the ways listed in the TMA, relying on arguments such as that "affixing" requires active behaviour. It says that any use that is deemed to be in scope should be of a similar ilk to that listed in the TMA, which is not the case here. 

Generally, the courts have taken a fairly expansive view of the types of use that can constitute infringement. A good example is the recent Samsung case (not relied upon by either party), where the Court of Appeal confirmed that there was use even though the type of use could not necessarily be specified. Indeed, the CJEU has gone as far as saying that no use (de-branding) can qualify as infringement. Getty will be hoping that this trend continues.

Is use by Stability?

Stability argues that any use is by the end user, not Stability, relying on case law drawn from the intermediary liability and keyword advertising contexts. This case law says that use involves "active behaviour" and "direct or indirect control" of the act constituting the use. It also implies use in the party's "own commercial communication".

Stability contends that it has no "control" over the production of watermarked images and that such images are not "statements of commercial communication" as they are not real images. It says that merely creating the technical conditions in which use can occur is not enough and points to its terms of use, which put responsibility for image generation and permitted uses with the user.

Getty says that the cases relied upon by Stability all represent exceptions to the general principle that use by a party in the course of trade constitutes use by that party. It says that none of the case law cited by Stability is relevant given Stability's active role in training the models on images bearing the Getty marks. It argues that Stability must be legally responsible for watermarked outputs as they are produced by Stability's model without users specifically prompting for them and (as Stability itself says) users themselves do not want watermarks.

It will be interesting to see whether the judge applies existing case law to this issue or addresses it from first principles. It will also be interesting to see whether the fact that Stability was involved in the training of its model affects its liability for alleged infringements committed during the deployment of the model.

Is Stability a person?

The parties haven't submitted any arguments on whether use of the marks is by a "person". That's probably right as the UK government's response to its Call for views on AI and IP in 2021 specifically said that "The government accepts the view provided by respondents that AI systems may infringe trade marks but that AI should be regarded as a tool being deployed under human direction and that liability lies with a legal person." Effectively, the word "person" in the TMA does not preclude liability where the infringement is mediated through AI; the legal person behind the AI will be liable.

Is use "in relation to" goods/services?

This is one of the most important issues in dispute. We know from the case law that use has to be "in relation to" goods/services, meaning that it must create a material link between the mark and the goods/services in question. This requirement has been further interpreted to mean that use must be "as a trade mark" – "to indicate the trade origin of the goods/services in question".

The net result is that the mere fact that a trade mark appears on a product does not automatically mean that such use is “in relation to” that product; there must be "trade mark" use. For example, the use of a trade mark on the kit of football players depicted on collectable football cards does not constitute trade mark use "in relation to" those cards (Trebor Basset (HC)).

In most infringement cases, this issue is not contested. The Getty case is different. Here:

• Stability argues that there is no "trade mark use" as the average consumer will view watermarked outputs for what they are – accidental artefacts of the AI training process, not indicators of trade origin.

• Getty argues that consumers will believe that the watermarks indicate trade origin. Either they will believe that a watermarked output originates from Getty or was generated by a model which was trained on Getty assets under licence from Getty. 

In deciding this issue the judge will have to come to a view on issues such as the extent to which the average consumer will dwell on the watermarks, how technically savvy the average consumer is, the impact of the fact that some of the watermarks are garbled, how much of the surrounding context (such as the overall branding of Stability's model and its terms of use) is relevant, the weight to be given to the small number of users who have commented on the watermarks 'in the wild' and the extent to which watermarked outputs are ever viewed by others 'post-generation' such that their views are relevant.

The parties' arguments on these issues are summarised in the rider at the end of this article. The important thing to note for now is that, if Stability wins on this issue, all of Getty's trade mark infringement claims should fail (since there will be no use of the relevant type for infringement). However, if Getty wins on the issue there is a good chance that it will also win on its likelihood of confusion claim: if the average consumer views watermarked outputs as indicators of trade origin then they are also likely to be confused as to Getty's involvement.

Given what a pivotal issue this is, it is quite possible that whatever conclusion the judge comes to will form the basis of an appeal by one or other party. Depending on the judge's exact reasoning, we might therefore see an appeal on the extent to which context is relevant for infringement (yet to be fully decided). We might also see an appeal on the requirement for use to constitute "trade mark use". Indeed, Kerly, a leading tome on trade mark law, questions the desirability of this requirement and notes how it cuts across other areas of trade mark law. Given that the Supreme Court has recently ruled in the Iconix case that damage to trade marks can occur in a variety of ways (without confusion at the point of purchase), now might be a good time for Getty to argue that damage can occur to its marks even if there is not "trade mark use".

However, it should be remembered that, if this case fails on this ground, it might not mean that all AI cases suffer the same fate. Whether there is "trade mark use" will vary depending on the particular facts of each case.

Is use in the course of trade?

Use “in the course of trade” requires use in the context of a commercial activity with a view to economic advantage and use that is not a private matter. As Getty notes in its Skeleton, it's not clear from the case law whether these are cumulative (different) requirements or two sides of the same coin. What is clear, however, is that the courts usually consider whether there is use "in the course of trade" at a relatively high level of generality.

Getty makes that point and argues that Stability is running a business in the UK and providing its model as part of that business, so both conditions are met. Conversely, Stability repeats its arguments that outputs cannot be commercial communications from Stability as it has no legal or practical control over them.

Is use without the consent of the trade mark owner?

Stability has pleaded that Getty consented to the use of the watermarks where Getty generated the watermarked outputs itself as part of the litigation. Given that many trap purchases are used as evidence in litigation, it would be surprising if the judge ruled out Getty's evidence on this basis. Indeed, the judge called this issue a "damp squib" at trial. Clearly, Stability has only raised this issue to reinforce its argument that most of the allegedly infringing outputs have been generated by Getty for the purposes of the litigation and don't reflect what would happen in real life (ie its wilful contrivance argument).

Is there wilful contrivance by Getty?

The main arguments on this relate to later versions of Stability's model since Stability has admitted that the incidence of watermarks is "non-trivial" for earlier versions. For later versions, Stability says that the incidence is "vanishingly rare" and criticises the prompts used by Getty to generate watermarked outputs as contrived.

Getty argues that Stability has not run a de minimis defence and that infringement is not contrived as a matter of fact since there is evidence that watermarked outputs have been generated by later versions of Stability's model in real life by end users using a variety of prompts.

While it looks as though Stability has had some success in filtering out watermarks for later versions of its model, it doesn't look like it has been successful (yet) in removing them entirely. However, neither party was able to put forward convincing evidence at trial on the probability of a watermarked output being produced by later versions of the model. It is therefore difficult to predict where the judge will come down on this issue.

Assuming the judge accepts Stability's argument that the likelihood of a real world infringement for later versions of its model is very low to nil, she will have to consider what this means: is there no infringement, is it reflected in the remedies, or is it irrelevant? There is no precedent for this in the UK. Any findings the judge makes on this might also be relevant to future copyright infringement claims where the incidence of infringing outputs could also be in issue (as it was in this case before the relevant copyright claim was dropped by Getty).

Who is the average consumer?

The parties disagree on who the average consumer is and how technically competent they are. Stability argues that the average consumer is technically competent, particularly for the downloadable and developer versions of its model, which require some skill to download and use. Getty argues that the average consumer is the public at large and not particularly technically savvy. It says that users will ask for help from more technically competent friends and family if they cannot download or operate models themselves, just as some witnesses did for this case.

There is an interesting possible outcome here: there could be a finding of no infringement for the downloadable and developer versions of the model (where the average consumer is technically savvy and sees the watermarks for what they are) but infringement for the cloud-based version of the model (where the average consumer is the public at large and sees the watermarks as confusing indicators of trade origin). This all raises the question of whether the risk of infringement could change over time as users become more technically competent with AI and the factual matrix alters!

What level of attention will the average consumer pay?

Getty says that the average consumer would pay a moderate degree of attention. Unsurprisingly, this flexed during the trial to moderate to low. Stability is seemingly silent on the point, at least in its written submissions.

Is there direct infringement under section 10(1) of the TMA?

This requires the use of an identical mark in relation to identical goods. There is also a requirement (not apparent on the face of the legislation) that the use affects one of the essential functions of the trade mark.

• Identity of marks. Some of the allegedly infringing watermarks are garbled, which is the only difference between them and Getty's registrations. At trial, the judge seemed willing to accept that some watermarks might be identical to Getty's marks, but questioned whether there would come a point where a watermark is so garbled that it is not identical. It is therefore possible that the judge will make a finding of identity for some allegedly infringing watermarks but not others.

• Identity of goods/services. This comes down to whether or not watermarked outputs can properly be described as "photographs" or "digital illustrations" - items within Getty's specification. Getty says they are; Stability says that its outputs are never described like that and that it is offering a service, not goods. 

• Essential function affected. This issue boils down to whether or not the allegedly infringing watermarks are being used as (confusing) indicators of trade origin or not (the same issue as discussed above). Getty has not argued that any of the other functions of its mark are affected as its entire case is predicated on origin confusion.

Given what she has said at trial, it is possible that the judge will find identity of marks and goods/services for at least some of the outputs. If so, whether the specific elements of this ground are made out might come down to whether or not there is (confusing) "trade mark use".

Is there a likelihood of confusion under section 10(2) of the TMA?

Since similarity of marks and goods/services is admitted, this ground turns on whether there is a likelihood of confusion. Here, the parties substantially rehash their arguments as to whether Stability's use is "trade mark use", as already summarised above. It seems likely that whichever party wins on the "trade mark use" point will also win on a likelihood of confusion since the same issues are in dispute.

Is there unfair advantage/detriment under section 10(3) of the TMA?

Stability admits the existence of the required reputation and link and does not really contest the "without due cause" element. These claims therefore turn on whether or not there is a change in the economic behaviour of consumers (or a serious risk of it) and one of the specified types of injury.

• Change in economic behaviour. Getty's main argument is that there is a change in the economic behaviour of consumers as they might use Stability's model to generate infringing images to avoid paying Getty's licence fees. This seems like a difficult argument on the facts of this case. If the judge rejects it, then all of Getty's section 10(3) claims could fail, assuming that she also holds that a change in economic behaviour of consumers is a requirement for all types of injury. If she agrees with Getty that it is only a requirement for dilution, then Getty might still win on tarnishment and unfair advantage. There could be an appeal on this point. 

• Dilution. The difficulty with this ground is that even Getty acknowledges that there must be a change in the economic behaviour of consumers for it to be made out, which seems difficult on the facts.

• Tarnishment. Getty's claim is that any use of a watermark on outputs featuring pornography, violent imagery and/or propaganda would tarnish the Getty marks. At trial, the judge said that she could accept that a watermark and unsavoury content might be produced, but that she had difficulty accepting that both would occur in the same output. Getty argues that it has already happened – that watermarked images generated of Miley Cyrus constitute pornography and of Donald Trump constitute propaganda. This claim might therefore turn on whether the judge accepts this argument - and that tarnishment is likely to result. Some view this as a stretch on these particular facts, but the Court of Appeal interpreted the word pornography broadly in its recent judgment, so let's see.

• Unfair advantage. Stability says that it obtains no advantage from the use of the watermarks since they make the outputs unsuitable for use – that's why it is seeking to filter them out. It says that it lacks an intention to take unfair advantage, which is often a key guide to whether or not there is unfair advantage in practice. Getty says that Stability gains an unfair advantage (whether intended or not) as the watermarks suggest that Stability has the backing of Getty ie that there is a simple transfer of image. Again, this might come down to whether the average consumer views the watermarks as indicators of trade origin (such that those outputs appear to have the "backing" of Getty) or mere artefacts of the AI training process! 

Overall, Getty's section 10(3) claims seem more finely balanced than its direct infringement and likelihood of confusion claims. 

What does this mean for you?

• The case is legally and factually complex. A number of issues turn on the question of whether the average consumer views the watermarks as (confusing) indicators of trade origin or merely as artefacts of the AI training process. 

• If the judge holds that the average consumer views the watermarks as mere artefacts of the AI training process (such that Stability 'wins' the case), it does not necessarily mean that all AI cases will fail. This case involves the unauthorised use of a watermark, which is not necessarily comparable to other conceivable unauthorised uses. That's not to say that all aspects of the ruling will be irrelevant to other disputes. The rulings on whether the Getty marks are being "used", whether that use is by Stability, whether it is in the course of trade and the like will all be important and of potentially wider application. We might also have some interesting rulings (and appeals) on the relevance of context to infringement, whether there must be a change in economic behaviour for all reputation-based claims, the meaning of the "in relation" to wording and how existing law on actual confusion and post-sale context should be applied. 

• If the judge holds that the average consumer will view the watermarks as indicators of trade origin, then she will probably also find a likelihood of confusion. Infringement will then turn on the judge's findings on all of the other general aspects of infringement discussed at the start of this article. 

• If Getty wins on infringement, it might only win in relation to certain versions of Stability's model (eg the earlier versions of the model for which watermarks are more common or the cloud-based versions where the average consumer is the less technically savvy public at large). The former would 'blunt' any win. 

• Any victory for Getty would be somewhat pyrrhic if Stability is able to filter out use of the Getty watermarks in future. Others who use AI might not want to – or be able to – undertake such filtering, which is another significant point of difference between this dispute and other potential AI disputes. 

• There are various additional allegations from both sides not dealt with in this article such as ineffective pleading by the other, that Stability lacks responsibility for particular models and that certain models are not infringing at all. All of these arguments could impact the outcome of the case, as could some of the confidential material not addressed in open court.

• A decision is expected in the autumn term. It is likely to be one of the most interesting trade mark infringement decisions in a long time, covering (as it does) almost every element of infringement. However, it's unlikely to be the start and end of such cases. There is still much to be worked out. As we see more entities developing and deploying AI, there are likely to be ever more complex issues around Louboutin-type use, contributory infringement and more.

  ---

Rider: the parties' key arguments on whether watermarked outputs would be viewed as (confusing) indications of trade origin or mere artefacts of the AI training process

Getty argues that:

• The average consumer will view the watermarks as denoting the trade origin of the outputs. Either they will believe that a watermarked output originates from Getty or was generated by a model which was trained on Getty assets under licence from Getty.

• The average consumer is the public at large. They are not particularly technically savvy and will only pay a moderate degree of attention (adjusted to moderate to low at trial). They won't dwell on the watermarks. Stability is wrong that the average consumer of the downloadable and developer versions of its model is a technically competent sub-group of the public. If downloading these models is tricky, users will ask technically competent friends or family to undertake the download for them. 

• The use of a watermark is the epitome of use to indicate trade origin. The Getty watermarks generally appear in the same location and manner on Stability's outputs as on Getty's assets. Users are used to seeing Getty watermarks, which are often faint, on a variety of content. 

• It is not appropriate to take an overly-expansive view of context when deciding infringement. Matter which is not visible to the average consumer when it is using the mark (such as terms of service) is irrelevant. The fact that Stability's models are branded with Stability's marks does not dispel confusion as to trade origin, especially as there is no evidence that Stability's marks are distinctive. 

• There is at least one example of a person wondering if Getty has a relationship with Stability. While Reddit exchanges indicate that a small number of users understand why watermarks are generated, other real world users are left "scratching their heads". The fact that there is not more evidence of actual confusion is not decisive and hardly surprising given the ephemeral nature of the outputs. 

• The views of those who encounter a watermarked image post-generation also need to be taken into account, especially following the recent Iconix Supreme Court ruling. It is entirely realistic that some consumers will be shown another’s screen or image bearing the watermark without having been involved in the preparatory set-up process or the generation of the image.

Stability argues that:

• Users will view the watermarks for what they are – accidental artefacts of the training process, and not real. Watermarks make the images unusable and their presence indicates that an image is unlicensed.

• Users are technically savvy and know how AI is trained and works. This is particularly so for users of the downloadable and developer versions of the model, which require some technical competency to install and operate. 

• Consumers have expectations about how brand messaging is deployed. The watermarks do not conform to those expectations – evident by the way the watermarks are presented (including some being garbled), the overall branding of the model, the lack of any indication of a co-branding arrangement and the type of images produced (which are not the type normally promulgated by Getty). 

• Stability's terms of service support all of the above.

• The example of actual confusion put forward by Getty involves a person in the USA (not the UK), who, in any event, does not think that there is a relationship between the parties. Reddit and other exchanges show that the average consumer knows that the watermarks are unwanted artefacts of the AI training process. Getty would have uncovered more evidence of actual confusion if it existed as Getty has left "no stone unturned". 

• Getty has not correctly pleaded the post-generation argument and (in any event) those downstream of the user will not encounter watermarked outputs as the user will immediately discard them as useless and generate new images.