When the overlap of AI and trade mark law is discussed, it is often relating to ensuring good clearance practices are adopted in relation to branding generated by AI. While that remains an important consideration, the liability for the appearance of marks in generative AI outputs poses an underappreciated risk for AI platforms and intermediaries following the recent Getty v Stability AI decision.

For all trade mark infringement, there must be use, by the defendant, of a sign, in the course of trade. The Trade Marks Act 1994 (‘TMA’) offers various scenarios that fulfil this criteria at s10(4) TMA, none of which clearly apply to an output. Historically, platforms have been found not to be using trade marks in providing their services; most famously, Google avoided being held liable for the appearance of trade marks in its sponsored advertisements (in the Google France case). There have been notable exceptions to this general rule, particularly the case of Swatch v Samsung, where Samsung was found liable for infringing watch faces made available through its platform on the basis that it had too much involvement in the marketing of them.

Further, there may be other circumstances where a business will have made the technical arrangements necessary for a trade mark infringement to occur, and despite having had no oversight of the content itself, it could be liable. A significant question would remain over whether the provider is using the mark in a way that will affect the functions of a trade mark (a criteria necessary for infringement).

Case law has developed which explains that the use of the sign must be the defendant’s own ‘commercial communication’ for them to be held directly liable. To assess whether use is a party’s commercial communication there needs to be sufficiently ‘active’ involvement from the perspective of the average consumer such that “a well-informed and reasonably observant user might believe that the operator is marketing, in its own name and on its own account, the goods for which the sign in question is being used”¹.

In Getty the analysis of active involvement of Stability in the production of outputs that contained signs similar to Getty watermarks was an issue before the court. Stability’s position was that the user was responsible for the signs appearing in generative AI outputs because the user was responsible for the prompts. They argued that Stability’s involvement was passive and therefore they were not using the signs in the course of trade. Both Getty and the judge disagreed:

• Stability had control over the training materials which included third party IP;
• they had the ability to flag certain words at the prompt stage that would divert the standard output process (i.e. prompts intending to create inappropriate or dangerous outputs); and
• following the prompt stage the output could likewise be edited to filter inappropriate outputs.
• In contrast, the user had no control over these aspects. This level of control by Stability was more than was the case in previous cases such as Google France.

An additional consideration for whether use is in the course of trade is the perception of the average consumer. Whilst this is a key point for likelihood of confusion for infringement under s10(2) TMA, the consumer perspective also comes up clearly in the commercial communication test. In Getty, Stability argued that consumers would not believe the platform was responsible. Some of the key reasons for this view was the contention that consumers consider themselves to be responsible for the output; they would see the inclusion of any signs similar to watermarks as a technical result rather than an intended badge of origin; and the absence of specific co-branding would indicate there was no commercial connection between Getty and Stability. For similar reasons that Stability was considered to control the outputs, it was found that users would not consider themselves responsible for outputs (particularly in circumstances where the users did not want watermarked images) and would assume that, where the watermarks were relatively undistorted, Stability had some form of consent to use their trade marked works, or association with Getty.

Future battlegrounds for trade mark liability?

Generative AI tools are a completely different scenario to keyword search advertising in the 00s. In particular, ‘chatbots’ output text that the user may have no idea what it will say until it says it – often that is the reason the user has entered the prompt. The question is: when the AI tool includes a trade mark in its answer or output, is it infringing a trade mark by doing so?

Two sectors we think are prospects for future litigation are advertising and publishing:

An interesting frontier for potential dispute is the overlap between product recommendations and advertising. If one asks an AI chatbot for recommendations for a product type, it will likely include trade marks in its response. Indeed, it would not be especially helpful if it did not include Nike, adidas, ON etc. in a response to a question about which running shoes to buy. Most product-led brands’ trade marks will not have coverage for AI outputs; they will be focused on their product categories (e.g. ‘trainers’ in the earlier example), meaning that strict liability ‘double identity’ infringement (same sign, same goods/services) will be off the table, and it is unclear how, without some underlying commercial incentivisation between brands and the AI platforms, there would be any confusion, unfair advantage or detriment to the trade mark. Any trade mark use may also be sufficiently editorial (or ‘referential’) to fall within the ‘honest commercial practices’ exception to infringement. However, as advertisers look to move into the environments that are replacing the search engine results page, this may change.

Another industry to watch is publishing. AI platforms often reformulate and republish content from online publishers. A chat that summarises the source publication might be considered to be directly competing for the publisher’s market. The summaries may also not be accurate given the current limitations of the technology powering large language models. Where an AI output credits their source, as is becoming more common, is the platform using the trade mark in the course of trade and in its own commercial communication? Is such use ‘in accordance with honest commercial practices’ where the content may have been scraped without permission from the publisher’s website. This all points towards increasing licensing of publisher content.

In the meantime, the Getty v Stability case is heading to the Court of Appeal, so watch this space!

Stay tuned for more from our series of insights as we count down to INTA 2026. For questions on AI and trade mark infringement liability, or to arrange a meeting at INTA, contact Oliver Fairhurst or Victoria Owrid.