Getty Images v Stability AI: the trade mark ruling

Getty won on double identity infringement and likelihood of confusion, but only for certain Stability legacy models. It failed on its reputation-based claim. It means Stability should be free to continue to supply its current models in the UK.

Here, we look at why Getty's trade mark infringement claims were only partially successful and what it could mean for future claims. For the background to the case and the allegedly infringing outputs, see here. 

Why were only certain versions held to infringe?

It was common ground that Getty's trade mark infringement claims had to be considered separately for each version of Stability's model. This was largely because "Models have not all been trained on the same dataset and different filters have been applied to the training data. Thus while one version of the Model may generate synthetic images which include the Sign, another version may not." 

The judge found various reasons why certain versions of Stability's model did not infringe, including that:

• Stability was not responsible for an earlier version of the model released on certain platforms. 

• Later versions of the model had never been shown by Getty to produce even one output containing a Getty watermark.

• Some of the watermarked outputs generated by users 'in the wild' could not be linked to a specific version of the model.

• At least one watermarked output was generated by a pre-release (beta) version of the model, which underwent substantial 'de-watermarking' prior to release. 

This was all supported by other evidence before the court, which suggested that earlier versions of Stability's model had a greater propensity to produce watermarks, with filtering becoming more successful as time went on. It was also supported by Stability's admission (which it later tried to pull back on) that the incidence of watermarked outputs was 'non-trivial' for earlier versions of its model. The net result was that only legacy versions (v1.x and v2.x) potentially infringed, not later models (SD XL and v1.6). 

Key learning: In most cases, it should be relatively easy for brand owners to show that the current version of a model infringes since there is unlikely to be filtering of the type seen in this case. Nonetheless, brand owners will need to be alert to potential arguments that only some - and not other - versions of a model infringe and will need to clearly link allegedly infringing outputs to specific model versions. 

Why were certain outputs disregarded: prompts?

The judge also disregarded certain watermarked outputs on the basis that the prompts used to generate them were contrived and did not reflect real-world usage. Users of Stability's model would not copy Getty captions, use re-worded captions or derivatives of captions. They might use words such as 'vector art' and 'news photo' though – indeed, there was (limited) evidence that some users had.

Interestingly, however, the judge did not disregard certain watermarked outputs solely because it was not clear whether they had been generated in the UK, seeing "no difference… between users in the UK and elsewhere". 

Key learnings: Brand owners must make sure that the prompts used to generate allegedly infringing outputs are realistic and have not been contrived to 'force' infringement. Specific evidence on the legitimacy of particular prompts might be needed. Real-world evidence of infringing outputs will obviously carry more weight. Whether non-UK outputs are probative would depend on the particular model versions used across the world.

Why was infringement considered on an output-by-output basis?

The judge held that "Unlike the more usual case of trade mark infringement, it is impossible to conclude that for every watermark* generated by the same version of the Model, a similar analysis will apply such that infringement would follow automatically." For example, a watermark might be more or less garbled, impacting whether there is infringement. 

Key learning: Unless it can be shown that the presentation/use of the allegedly infringing mark is fairly uniform on outputs, brand owners may have to argue infringement on an output-by-output basis. In other words, it might be difficult for the court to extrapolate infringement based on one or two examples. This weighs in favour of claimants not overly tying their infringement claims to a small number of representative infringements, which seemed to undo Getty here. 

What level of infringement was required: de minimis?

Stability eventually conceded that it had not pleaded a de minimis defence so the judge did not have to consider this issue. It was therefore agreed that Getty need only establish "on the balance of probabilities, that each version of the Model has generated at least one output with a watermark…" in the UK. 

It is disappointing that a de minimis argument was not run, as this issue has been little considered by the UK courts (the Merck case being a notable exception). Key learning: Brand owners will need to consider de minimis-type arguments carefully where allegedly infringing outputs are rare. It could impact liability and/or remedies, but the paucity of case law on the point makes it difficult to predict that impact and the answer is likely to be fact-specific. Evidence of the probability of an infringing output (which the parties did not produce in this case) is likely to be important, assuming it can even be calculated.  

What was the relevant date for infringement?

The judge held that "It is impossible to identify when each Model will have generated an allegedly infringing watermark, and so I accept Getty Images’ case that the relevant dates for present purposes must assume infringement from the date of release of each Model."

Had Stability 'used' the watermarks in the course of trade etc?

All acts of infringement, irrespective of the particular section under which they arise, require 'use' of the mark by the claimant in the 'course of trade' 'in relation to' goods/services. The judge held that each of these preconditions for infringement was present:

• Stability (and not end users) had 'affixed' the watermarks to the outputs and offered and put outputs on the market. 

• While users had some control over the outputs, Stability had the most control. Watermarks only appeared because of the training and making available of the particular model by Stability; users did not want the watermarks and did not prompt for them. 

• Since Stability runs a business in the UK and offers its model as part of that business, use was in the course of trade.

• Use was 'in relation to' goods/services (ie origin-identifying use, as discussed here) – this is discussed below in the context of likelihood of confusion.

Key learnings: The decision builds on the body of case law on 'use' established in the intermediary liability and keyword advertising cases. While brand owners will be relieved that Stability (and not the end user) was deemed to have 'used' Getty's mark, it's unclear from the judgment whether the same would be true where the end user specifically prompts for a particular mark. Brand owners will need to consider this aspect of the ruling carefully. 

Who is the average consumer?

By closing submissions, it was common ground that there were at least three categories of average consumer in this case, comprising those who use (a) the downloadable model, (b) the developer model and (c) the web-based model. The degree of technical skill would be graduated with (a) being the most technically savvy and (c) the least. Such consumers would pay a moderate and perhaps even high degree of attention. 

Key learnings: Brand owners should take note that the average consumer might vary depending on how the particular AI model in question is accessed/used and the other facts of the case. This might, in turn, impact the assessment of infringement. 

Was there double identity infringement (s.10(1))?

The judge held that the watermark on the left below was identical to Getty's word marks (even though there was some blurring), while the mark on the right below was not (because it contains an extra 'I' in images). There is no new law here – marks are identical if the differences are so insignificant that they may go unnoticed by the average consumer. However, this is an interesting application of the law.

Identical	Not Identical

The judge held that Stability was using the watermarks on synthetic image outputs – a use identical to certain items in Getty's trade mark specifications (including digital imaging services, downloadable digital illustrations and graphics, and digital media). It meant that there was s.10(1) infringement for iStock watermarks but only those generated by earlier versions (v1.x) of the model.