Decreased confidence in loss-avoidance contexts is a primary meta-cognitive bias of human reinforcement learning

Abstract

In simple probabilistic instrumental-learning tasks, humans learn to seek reward and to avoid punishment equally well. Despite this remarkable symmetry in choice accuracy between gain and loss contexts, two recent effects of valence have been independently documented in reinforcement learning. First, decisions in a loss-context are slower, which is consistent with the Pavlovian-instrumental transfer hypothesis. Second, the loss context decreases individuals’ confidence in their choices. Whether these two effects are two facets of a single process or two independent effects of valence is still unknown. Here, across five experiments, we assessed the relative merit of the two hypotheses. Our results show that, in loss-contexts, the decrease in confidence in one’s choices can be robustly observed in the absence of the response time bias. This suggests that the effects of valence on motor and metacognitive responses, although concomitant in most cases, are dissociable. Jointly, these results highlight new important constraints that should be incorporated in mechanistic models of decision-making that integrate choice, reaction times and confidence.