Exploring Personality Structure Through LLM Agent

Abstract: Lexical studies seek universal personality dimensions by analyzing trait-descriptive words. This study examined whether large language model (LLM) agents, endowed with Big Five profiles, could reproduce human lexical structures. GPT-4o agents rated Japanese trait words, and principal component analyses with Bass-Ackward comparisons were conducted. The broadest dimensions, corresponding to the Big Two, were robustly recovered, supporting their cross-linguistic universality. Four stable components also emerged, including prosociality and conscientiousness, whereas activity/introversion and Neuroticism appeared less consistently. Compared with human data, LLM responses showed higher interitem correlations, greater internal consistency, and an exceptionally large first component, suggesting blurred distinctions between traits. Congruence with human data was lower than human–human benchmarks, reflecting multilingual training influences and biases. These findings indicate that LLMs can recover broad dimensions but impose structural compression. Future research should test prompt variations, alternative models, and baseline conditions to clarify sources of LLM-generated structures.