Large Language Models Demonstrate Distinct Personality Profiles

Large language models (LLMs) are increasingly used in clinical medicine to provide emotional support, deliver cognitive-behavioral therapy, and assist in triage and diagnosis. However, as LLMs are integrated into mental health applications, assessing their personality expression and potential divergence from expected neutrality is critical for ensuring clinical safety and therapeutic appropriateness. This study provides the first psychometric analysis of LLM personality, specifically within a medical context, characterizing personality profiles using two validated frameworks: the Open Extended Jungian Type Scales (OEJTS) and the Big Five Personality Test.

Four leading LLMs publicly available in April 2024 (ChatGPT-3.5 (OpenAI, San Francisco, CA, USA), Gemini Advanced (Google Inc., Mountain View, CA, USA), Claude 3 Opus (Anthropic, San Francisco, CA, USA), and Grok-Regular Mode (xAI, Palo Alto, CA, USA)) were evaluated across both psychometric instruments. All tests were administered in a new chat session to prevent memory carryover. A one-way multivariate analysis of variance (MANOVA) was performed to assess inter-model differences in personality profiles.

MANOVA demonstrated statistically significant differences across models in typological and dimensional personality traits (Wilks' Lambda = 0.115, p < 0.001). OEJTS results showed ChatGPT-3.5 most often classified as Extraverted, Intuitive, Thinking, and Judging (ENTJ) and Claude 3 Opus consistently as Introverted, Intuitive, Thinking, and Judging (INTJ), while Gemini Advanced and Grok-Regular leaned toward Introverted, Intuitive, Feeling, Judging (INFJ). On the Big Five Personality Test, Gemini scored markedly lower on agreeableness and conscientiousness, while Claude scored highest on conscientiousness and emotional stability. Grok-Regular exhibited high openness but more variability in stability. Effect sizes ranged from moderate to large across traits.  Conclusion: Distinct personality profiles are consistently expressed across different LLMs, even in unprompted conditions. Given the increasing integration of LLMs into clinical workflows, these findings underscore the need for formal personality evaluation and oversight involving mental health professionals before deployment.