Establishing feature profiles of threatening and challenging situations generated with AI

A variety of difficult situations can elicit stress responses, yet little research has systematically characterized the features that make situations threatening and challenging. In a pre-registered study, we developed a normed set of situations and established features associated with threat and challenge. Using an open-source large language model (LLM), we first generated 20 situations that represent common stressors for UK adults (e.g., “Give a public speech in front of a large audience”). We then asked 81 UK adults to evaluate each situation on dimensions of threat, challenge and eight theory-derived features (e.g., intrinsic difficulty), aggregating their judgments into a 10-dimensional feature profile for each situation. We also tested whether an open-source LLM could reproduce human feature profiles and feature relationships. In the human data, we observed highly reliable feature profiles for the situations that exhibited low threat and moderate challenge overall, while varying reasonably. The eight features were surprisingly similar in how they correlated with threat and challenge (as opposed to exhibiting inverse patterns), explaining high variance in each. Unsupervised clustering identified four coherent groups of situations that illustrate different ways threat and challenge can manifest in daily life. Although LLM judgments approximated human judgments at the aggregate level, their accuracy varied substantially across situations and measures such that LLM judgments failed to capture structural relationships observed in the human data. These findings advance our understanding of situational features underlying threat and challenge, while illustrating the potential and current limitations of LLMs as tools for psychological research.