Human outcome-irrelevant value learning is insensitive to its monetary cost

Abstract

When learning from reinforcement, individuals should assign credit selectively to features of their actions that causally determine outcomes. Yet, even when explicit instructions and extensive experience clarify which action features predict reward, humans assign credit to outcome-irrelevant features, leading to suboptimal choices. Here, we use computational modeling to test if this suboptimal credit assignment can be controlled depending on the monetary cost at stake. Specifically, we tested whether higher payoffs can increase control over outcome-irrelevant credit assignment. Participants (N = 287) completed a multi-armed bandit task under low and high monetary stakes. Higher stakes increased the influence of outcome-relevant values on choice, but, strikingly, had no influence on outcome-irrelevant values, which influenced choices to a similar degree. These findings demonstrate the automatic nature of credit assignment to outcome-irrelevant features, suggesting that associations between features and reward are formed despite clear instructions and their detrimental consequences for monetary gains.