Abstract

Humans learn to trust new partners by evaluating the outcomes of repeated interpersonal interactions. However, available prior information concerning the reputation of these partners may alter the way in which outcomes affect learning. This thesis combines for the first time behavioral, computational, psychophysiological and neural models in a direct comparison of interaction-based and prior-based decision-to-trust mechanisms. Three studies are presented, in which participants played repeated and single trust games with anonymous counterparts. We manipulated several conditions: whether or not reputational priors were provided, the probability of reciprocation (trustworthiness) of each counterpart, and the time-horizon of the relationships. The thesis addresses several challenges involved in understanding the complex behavior of people in social contexts, by investigating whether and how they integrate reputation into decisions to trust unfamiliar others, by designing ways to combine reputation information and observed trustworthiness into unified models, and by providing insight into information on the brain processes underlying social cognition. Numerous models, algorithms, game theoretical and neuroscientific methods are used to examine these questions. The thesis presents several new reinforcement learning (RL) models and explores how well these models explain the behavioral and neural interactions between trust and reputation. The performance of the new models was tested using experiments of varying complexity. These experiments showed that model-based algorithms correlate better with behavioral and neural responses than model-free RL algorithms. More specifically, when no prior information was available our results were consistent with previous studies in reporting the neural detection of parametric estimates of RL models within the bilateral caudate nuclei. However, our work additionally showed that this correlation was modified when reputational priors on counterparts are provided. Indeed participants continued to rely on priors even when experience shed doubt on their accuracy. Notably, violations of trust from counterparts with high pro-social reputations elicited both stronger electrodermal responses and caudate deactivations when priors were available than when they were not. However, tolerance to such violations appeared to be mediated by priors-enhanced connectivity between the caudate nucleus and ventrolateral prefrontal cortex, which was anti-correlated with retaliation rates. Moreover, in addition to affecting learning mechanisms, violation of trust clearly influenced emotional arousal and increased subsequent recognition of partners who had betrayed trust.