Predicting Personality Traits from Short Unstructured Conversations with AI

Given the complexity of personality, we might assume that significant behaviour is required to express our traits. Here we use large language models (LLMs) to test the extent to which personality is visible in short conversations with artificial intelligence (AI). In a preregistered experiment, 100 participants completed the Big-Five Inventory (BFI) before having unstructured conversations with GPT-4.1, writing a median of 430 words. We then instructed a second LLM rater (Claude Sonnet 4.6) to complete the BFI for each participant, predicting their responses based solely on an analysis of the participant-GPT chats. Correlations between LLM-predicted and self-reported BFI traits were significant (range r = .25 to .47), in some cases matching previously documented self-other predictions from family members. A linguistic analysis revealed that the LLM rater’s predictions relied largely on lexical content as opposed to syntax. Using preexisting data, we tested if the AI’s persona influenced how much participants revealed about themselves in conversation. A hundred participants chatted with two versions of GPT-4.1—one that mirrored their personality (mirror AI) and one that opposed it (inverse AI). The LLM rater predicted BFI traits based on the interactions and we correlated this prediction with self-report. Mirror chats generated significant correlations, averaging r = .41; for inverse chats the average fell to r = .30, dropping to zero when the prediction relied on lexical content alone. People readily convey their personality profile in short conversations with AI, and this is enhanced when the AI’s persona mirrors the users.