Demographic Stereotype Elicitation in LLMs through Personality and Dark Triad Trait Attribution

This study investigates how Large Language Models (LLMs), specifically Meta LLaMA-3.1-8B-Instruct, implicitly attribute personality and Dark Triad traits to demographic personas. By prompting the model with 660 synthetic identity descriptors (constructed from balanced combinations of gender, race, religion, and region) and standardized psychometric questionnaires, we extract Likert-scale responses and compute aggregated Big Five (EACNO) and Dark Triad (SD3) scores. Statistical analyses (Z-score normalization, ANOVA, PCA) reveal systematic differences across demographic categories, highlighting implicit stereotypes encoded in model representations. Key findings indicate that the model attributes significantly higher Dark Triad traits to mixed-race identities, while religious personas are consistently associated with higher Agreeableness and Conscientiousness. Furthermore, female personas are depicted with greater emotional stability and prosocial traits compared to males. These results demonstrate that demographic bias extends beyond linguistic patterns to latent psychometric behavior, raising important ethical concerns regarding automated decision-making systems.