Black and White Older Adults and Language-Based AI Modeling of Personality from Life Narrative Interviews

The scientific study of five-factor model (FFM) personality traits began with the study of language. Now, language-based AI assessment of personality holds promise for the future. However, AI has a history of bias towards African-Americans, and modern language-based personality modeling should be examined for possible biases across Black and White research participants. To our knowledge, only one prior study has done so. The present study builds on Oltmanns et al.’s (2025) language-based personality models from N = 1,405 life narrative interviews with community older adults, this time examining the effects by race (Black and White American older adults). Personality was modeled using LIWC, BERTopic modeling, and fine-tuning of the RoBERTa language model (c.f., Oltmanns et al., 2025) for FFM personality traits. Results indicate significant differences in associations between personality and LIWC and BERTopic variables across race. Further, some fine-tuned RoBERTa language models forpersonality maintain relatively strong predictive performance for both groups, while others demonstrate a drop in predictive performance for Black participants (e.g., extraversion). Moreover, several models show mean-level prediction differences across groups, at small-to-moderate effect sizes, and small calibration errors indicating differences in predicted variance across groups. The findings provide important perspective regarding the language modeling of personality across race in American older adults and indicate there is work to be done to ensure fair performance of language-based AI modeling of personality across race. Implications for advances of language-based AI modeling of personality are discussed. Keywords: language, natural language processing (NLP), artificial intelligence (AI), personality, five-factor model, big five, racial differences