Retailers Liability And AI Applications – Trademark

The question of whether the user (consumer) or the
provider of an AI application is liable for the infringement of
intellectual property rights as a result of the use of an
artificial intelligence application is an interesting one and
increasingly relevant to retailers as they roll out more and more
AI applications. The recent Getty Images vs Stability AI decision
of the 4th November 2025 has shed a bit more light on
this thorny issue.

Liability of intermediaries: an old story

The question of the level of liability of intermediaries in the
retail supply and promotion chain is not a new one. As referenced
in the Getty Images vs Stability AI decision, the Court of Justice
of the European Union (‘CJEU’) has handed down judgments
with regards the liability of search engine platforms such as
Google with regards keyword advertising in the well-known Google
France SARL and Google Inc. v Louis Vuitton Malletier SA
(C-236/08), with regards the storage and dispatch of goods in Coty
Germany GmbH v Amazon Services Europe Sàrl and others
(C‑567/18) and with the regards the use of trade marks by
third party advertisers in Daimler AG v Együd Garage
Gépjárműjavító és
Értékesítő Kft. (Case C‑179/15). In
all these cases, the intermediary was held not liable for trade
mark infringement based essentially on defences of lack of active
knowledge of any infringing activity and the platforms/parties
being essentially non-active in the infringing activity. In essence
the parties were held not to be using the trade marks concerned in
the course of trade, an essential pre-requisite for the finding of
trade mark infringement.

Now we come to the AI

Now, we come to the question of the liability of AI applications
which is a new phenomenon to consider. Stability AI attempted to
use the reasoning put forward in the Google France, Coty and
Daimler cases in the Getty case and argued that it was the user,
and the user alone, of its AI platforms who controlled the
potential output of infringing images created by them via the
prompts the user (consumer) entered into the AI applications.
Stability AI was essentially arguing it was not liable; it was
merely a tool not an active party in any infringing activity. Here
the court was considering the question of trade mark infringement
based on the production of the GETTY IMAGES and ISTOCK trade marks
in images produced by Stability AI applications. On this point,
Stability AI lost.

Counsel for Stability AI in cross examination with a witness
argued:

“the model is a tool controlled by the user and the more
detailed the prompt is, the more control is being
imposed.”

In response the witness responded:

“That is partially true. The user has control over what is
prompted. However, what the user does not have control over is what
the model is trained on. We have no control over that. What the
user does not have control over are any semantic guardrails that
might be put on the prompt and any semantic guardrails that we put
on the output. Absolutely, the user has control over what it asked
for but does not have 100% control over what is coming out the
other end.”

Stability AI lost this point because (a) it controlled the data
on which its AI applications were trained and (b) in the Getty case
at least the judge seemed to be swayed by the fact that most users
did not want GETTY IMAGES or ISTOCK trade mark appearing on their
outputted images, so it could be argued were actively not trying to
produce infringing images.

Now we come to the retailers and user’s
liability

The exchange between Stability AI and the witness quoted above
for me highlights some of the key issues of liability of AI
applications. A key issue seems to be control. It seems liability
cannot be pushed to the user, if the user has no absolute control
over the output of the AI application. Detailed prompts encouraging
infringing activity, such as searching for counterfeit products,
might make it more likely that liability could be pushed to user.
However, given that a user would rarely have complete control over
the outputs of an AI application, after all that is the point of an
AI application.

It seems that retailers might wish to concentrate more on
mitigating activities to counter infringing activities via their AI
application than trying to push the buck to the user. So here we
have to consider guardrails. It is interesting to note that
Stability AI had put in place so-called ‘filtering
functionality’ so the applications did not render
photo-realistic likenesses of well-known public figures, which were
added to combat concerns over fake news and propaganda. The prompts
were scanned for celebrity names. So, it was at least aware of the
potential of some concerning activity via its AI models.

So, what are the practical take aways for a retailer with an AI
application in place? For a retailer to avoid liability as best it
can and fall within the reasoning laid down in the Google France,
Coty and Daimler cases, it must in my eyes:

Screen the training data of the AI application to avoid
infringing material.

Put in place guardrails into the AI application to push users
away from infringing activity.

Put in place takedown procedures to remove infringing activity
when it is brought to its attention.

The above would push the retailer as much as possible into the
group of inactive enabler of commerce, akin to a keyword
advertising platform such as Google, not an active participant in
any infringing activity. Afterall the Google, Coty and Daimler
cases all concerned some arguably infringing activity, but those
parties succeeded in their defences because they were considered
not active in such activity.

What the Getty v Stability AI case however does however seem to
make clear is that once the AI application is deemed to using the
infringing sign in the course of trade liability will rest with the
owner/controller of the AI application, and that could in future be
a retailer.