¹Hong Kong Polytechnic University, Department of Applied Social Sciences, Hong Kong, Hong Kong, SAR of China

Background: Preventing harm to others is a foundational principle of human morality that relies on distinct learning processes when avoiding harm to others compared to oneself. Using a behavioral neuroendocrinology perspective, we investigated computational mechanisms underlying harm-avoidance learning.

Aims & Objectives: To investigate how testosterone administration affects learning to avoid harm in healthy men.

Method: One hundred and twenty healthy men randomly received either testosterone or placebo. Three hours later, participants completed a harm-avoidance task, learning over time to choose the option with a lower probability of electric shock to avoid harm either for themselves (Self condition) or for a stranger (Other condition).

Results: Behavioral analyses revealed that a quick catch-up of Other to Self existed in both groups, suggesting prosocial learning trends, but this catch-up tended to be delayed after testosterone administration. Computational modeling showed that a reinforcement learning model with separate learning rates for positive and negative prediction errors (PEs) best accounted for individuals’ choices. Consistent with the model-independent results, testosterone administration slowed down prosocial learning compared to placebo, particularly when people learned from negative but not from positive PEs. Testosterone also modulated the relationship between trait anxiety and prosocial learning from negative PE, specifying its anxiolytic effects.

Discussion & Conclusions: Together, our findings provide computationally precise insights into how testosterone administration contributes to prosocial learning and thereby affects harm avoidance for others.