In this talk I will discuss my experiences conducting role-playing exercises with larg
language models (LLMs), where I put artificial intelligence in the shoes of potential
study participants. I will cover two different use cases, one involving an LLM
interacting with a chatbot we designed, and one where I interviewed an LLM acting
as a hospital patient. In both cases, the aim was to identify points of communication
breakdown that could be addressed by improving the chatbot’s dialogue flow or the
interview questions before using the materials with human participants.
My experiences are generally positive, where these first trial runs with LLMs have
resulted in valuable insights to improve the study materials. This is especially the
case if the research team is small or in some ways different from the study
participants. Oftentimes, researchers have spent so much time with the materials
that they have developed blind spots that may be addressed by confronting them with
a “naive” LLM.
At the same time, it is important to touch upon several important critiques and
pitfalls. This includes the risk of overly relying on the LLMs’ responses and feedback,
which are subject to biases. This could create a form of “overfitting” by tailoring the
materials too much to LLMs’, instead of humans’, way of communicating.
Furthermore, the approach could be a tempting gateway to more extensive
involvement of LLMs in research, for example by considering the simulated interview
responses as usable and actionable data. Finally, it is challenging to justify the
prompt design for LLM-driven material testing, as there is not a clear set of quality
criteria for the LLM’s output in these contexts. This raises a critical question: by what
benchmarks can we determine if an LLM’s response is a faithful proxy for a human
participant?
Speaker: Jan de Wit