Large language models as psychometric laboratories? Maybe not yet. Cautionary evidence from synthetic personality data

This brief contribution compares human and GPT-generated responses to a short Big Five questionnaire to test whether LLM-simulated samples can stand in for human respondents in personality research. Synthetic data showed compressed variability, homogeneous profiles, and inter-trait correlations that diverged from the human structure, especially under persona prompting. LLMs may support item generation or exploratory testing, but should not currently replace human samples in psychometric validation.