Automatic value learning results in counterproductive human behavior

Humans are adept at accurately estimating the value of available choices from accumulated experience. However, cognitive processing also incorporates irrelevant information during deliberation, undermining decision accuracy. Here, we show that model-free credit assignment operates automatically, allowing irrelevant action features to acquire value and impair choice behavior. Across two preregistered studies (N = 504, 158) and a re-analysis of previously published data (N = 50), we rigorously examined outcome-irrelevant learning under conditions of explicit instruction and extensive training. In both cases, automatic value updating persisted even when participants clearly understood which features were non-predictive of reward, and reliance on these irrelevant values led to suboptimal choices. Moreover, working-memory capacity, known as a key regulatory resource, predicted the magnitude of this automatic outcome-irrelevant learning. Collectively, our results broaden the concept of automatic cognitive processing beyond the selective-attention literature to encompass reinforcement-learning mechanisms.