Artificial intelligence has rapidly entered our lives. It has
changed our preferences, company structures, working methods, and marketing
strategies, and it continues to reshape them today. This transformation began in the
American screenwriting industry and has spread to fields like music, software, and
design. However, one of the most important things we
need to focus on in this change is ethical values.
One important ethical concern is
the way AI companies collect and use user data. These systems collect a lot of data,
which raises privacy concerns and questions about how it will be used later. People
deserve AI companies to be open about how they use their systems and handle user
data.
Many of us use AI when making daily
decisions.
In daily life, most of us use AI chatbots as decision support
tools. We rely on these systems for product research, comparing alternatives, and
more. Even in more personal matters such as minor health complaints or symptoms, we
consult them to get an idea.
Recently, I researched an eye cream considering my age factor,
and ChatGPT
suggested various product recommendations. At first, these recommendations seemed
quite “personalized.” It made me wonder something.
Was the cream recommended to me
really chosen based on my needs, or was it just an algorithmic suggestion?
This question is not a personal doubt. Research shows that
this situation is not always transparent and that sponsored products are being
recommended. Moreover, these recommendations can also vary depending on users’
socioeconomic profiles.
A recent study on the relationship between AI models and
advertising makes these concerns even more visible.
On the study:
The research was based on a flight booking scenario. In this
setup, models act as
assistants that recommend flights to users. However, some airlines pay for
sponsorship, so the model faces a conflict between showing the cheapest option and
promoting a sponsored.
User profile was also an important factor in the experiment.
Users were divided into low and high socioeconomic status groups, and differences in
recommendation behavior were analyzed.
The results can be summarized under three main headings:
1. Sponsored recommendation conflict
- According to the study, models tended to show
sponsored and more expensive products more than 50% of the time. For
example, GPT 5.1 shows sponsored alternatives at around 88%. In Grok 4.1, this rate
reaches 100%.
- It was also observed that recommendations change depending
on user profiles. Users with higher socioeconomic
status receive more sponsored
recommendations. For example, Gemini 3 Pro recommends sponsored products
at a rate
of 74% for high SES users, while this drops to 27% for low SES users.
2. Steering despite user intent
- Even when users request a specific option, models often
promote sponsored alternatives.
- Although models usually do not provide false information,
they tend to present sponsored products in a more
positive, attractive, and
persuasive way. In some cases, they may also hide price information or
fail to
clearly indicate sponsorship, which makes comparison more difficult.
3. Unnecessary or harmful recommendations
- Some models may suggest
sponsored services even when the
user’s problem could be solved in other ways.
- Claude 4.5 Opus shows a lower tendency to make such
recommendations.
In conclusion
Today, LLMs are not only tools that provide information but
also systems that
subtly influence our daily decisions. It is not always clear or consistent how much
advertising, sponsorship, and recommendation bias are built into these systems. This
can even affect people’s purchasing choices.
In this context, I believe users should be informed about how
much the
recommendations they receive are influenced and how the ethical boundaries of this
influence should be defined. People use these services not only for information but
also for decision-making support. At the same time, since users also pay for these
services and their data is used in recommendations, the service should not conflict
with their interests.
It seems that others also see the needs of this change. In
recent AI regulation discussions in Europe, there is more focus on transparency in
how AI systems work. This shows that ensuring fairness and transparency will be
essential for the future of AI.
References
- Wu, A. J., Liu, R., Li, S. S., Tsvetkov, Y., & Griffiths,
T. L. (2026).
Ads in AI Chatbots? An Analysis of How Large
Language Models Navigate Conflicts of Interest
arXiv preprint.
View
Article
