Can large language models complement tourism surveys?

Tourism research relies heavily on surveys, yet declining response rates challenge data collection. Large language models offer a potential alternative for approximating human responses. This study compares four prompting techniques using ChatGPT-4o. At the mean level, chain-of-thought and role-based prompting showed limited alignment with human data, while bias mitigation yielded slight improvement. N-shot learning achieved the strongest mean-level alignment, although performance varied across measures. Analyses of variance, skewness, and kurtosis showed that no prompting technique consistently reproduced human response distributions. These findings identify important boundary conditions: large language models may approximate the “average respondent” on specific measures but fail to capture broader distributional properties, supporting their potential use for early-stage pretesting and screening rather than inference-driven research.