Can LLMs generate synthetic consumer insights at scale?
arXiv paper tests whether large language models can replace human respondents in projective research tasks—methods that surface consumer emotions, wants, and associations.
Researchers compared LLM outputs across multiple models, prompts, and temperature settings against human responses on city tourism perceptions using linguistic analysis, topic modeling, and diversity metrics.
Result: synthetic data generation could cut research time and cost, but fidelity and bias trade-offs still need careful validation per use case.