Author: Iványi Tamás (BME)
Abstract:
At first glance, the „knowledge base” of generative AI seems well suited to standing in for human research subjects, promising to accelerate the research process through synthetic data. A growing body of literature, however, warns that LLM-generated respondents come with systematic risks: reduced variance and homogenised answers, overly rational and socially desirable response patterns, stereotyped personas, and high sensitivity to prompt wording and model choice. Synthetic data that looks plausible is therefore not necessarily valid.
This presentation takes a methodological perspective and uses a case study from my own research. The case study illustrates which of the dangers described in the literature actually appear in practice, and which design choices help the researcher mitigate them: grounding personas in real sample characteristics, running repeated and varied generations, and validating synthetic results against human data before drawing conclusions. The presentation closes with practical recommendations on where synthetic respondents can already support marketing research, such as pretesting instruments and exploring scenarios, and where they cannot yet replace human participants.