Language AI enables endogenous topic discovery and behavioral prediction from trust game perceptions

The Trust Game (and its many variants) is a widely used experimental paradigm for studying trust and prosocial behavior, yet its interpretation remains debated. We demonstrate endogenous topic discovery and labeling via language AI as a scalable complement to traditional methods, uncovering semantic themes not commonly associated with the game. Using 1,124 subjects’ descriptions of Trust Game interactions from two studies, we show that social concerns emerge as key predictors of prosocial behavior, alongside trust-related language, even when the interaction is framed as a loan repayment. The pipeline converts unstructured responses into predictions of individual decisions: we train a cross-validated XGBoost–Random Forest ensemble that classifies prosocial and selfish behavior from topic features alone well above chance (AUC 0.73). Explicit language categorized as risk, gambling, investment, or game-theoretic themes was less informative in models predicting passing than language categorized as trust, teamwork, or involving moral self-image. These findings both validate and challenge conventional assumptions, highlighting the importance of social and moral language in managerial decision-making.