Digital twins of human social interactions reveal behavioral strategies in exchange games

Social interactions are an integral part of our daily life, and understanding their underlying mechanisms can reveal much about the cognitive and emotional processes that support human behavior, as well as their impact on overall well-being and mental health. Social interactions are predominantly studied using repeated social exchange games, such as the repeated trust game. In this context, behavior is analyzed in response to prede-fined computer strategies. This approach, however, can limit the generalizability of observed interaction pat-terns and hypothesized mechanisms. Recent advances in generative AI modeling have created opportunities to extrapolate data-generating processes and apply them to novel interaction contexts, even from short and noisy behavioral data. Such generative models allow studying social interaction behavior in virtual environments, where a model inferred from behavior can come to incarnate a digital twin of an individual’s interaction style. Through virtual simulations, these twins can then provide novel insights into underlying representations and mechanisms, as well as behavioral strategies. Here, we demonstrate the feasibility of this approach. We show that, from limited and noisy behavioral data obtained from a repeated trust game, it is possible to extract predic-tive behavioral generative dynamical systems models. We show that these models generalize to novel interac-tion settings and uncover behavioral mechanisms that extend beyond those observable in the empirical data alone.