Causal inference deficits and integration failure drive stable variation in human punishment sensitivity

Some individuals persist in behaviors that incur harm to themselves or others. While adaptive decision-making requires integrating such punishment feedback to update action selection, the mechanisms driving individual differences in this capacity remain unclear. Here, in a sample spanning 24 countries, we used a conditioned punishment task to identify how individuals learn from and adapt to punishment. We identified three, behaviorally robust phenotypes: (1) Sensitive, who correctly inferred punishment causality and adaptively updated decisions through direct experience of punishment; (2) Unaware, who failed to correctly infer punishment causality from direct experience but corrected their decisions following an informational intervention clarifying consequences; and (3) Compulsive, who persisted in harmful decisions despite both punishment and informational intervention. These phenotypes were driven by distinct cognitive mechanisms: (1) causal inference deficits, where individuals misinterpreted punishment causality, impairing correct knowledge acquisition (remediable via targeted informational intervention); and (2) integration failure, a deficit in synthesizing causal knowledge, action valuation, and action selection that rendered decision-making inert to punishment feedback, even after targeted informational intervention. Remarkably, these phenotypes predicted longitudinal outcomes (learning trajectories, choice behavior) six months later. By identifying the cognitive mechanisms driving variation in human punishment learning, this work provides a new framework to understand why individuals persist in harmful behavior and highlights the need for approaches to address these distinct cognitive barriers to adaptive decision-making.