Goals shape human confidence

Abstract

Confidence guides our decisions: it determines when we commit to a choice, gather more information, or change our minds. Prominent models treat confidence as a readout of the evidence in support of a decision: given the same evidence, a similar confidence level should follow, regardless of what the decision-maker is trying to achieve. Here we show instead that confidence formation is relative to the goals of the decision-maker. Across value-based and perceptual decisions in different sensory modalities, we manipulated the behavioural goal, requiring participants to select the option with either the most or the least evidence. Reversing the goal reversed the relationship between evidence and confidence: biases previously considered fixed features of a confidence computation, such as the positive evidence bias (PEB), followed the goal rather than the stimulus. We developed a new computational model, GOAL (Goal-Oriented Asymmetric Likelihood), in which evidence is weighted according to its relevance for the current goal, and validated it in behavioural and physiological (pupillometry) experiments. Our findings support a new account of metacognition in which confidence is constructed relative to the goals of the decision-maker, prioritising goal-supportive evidence.