Discrete confidence levels revealed by sequential decisions

Whether humans are optimal decision makers is still a debated issue in the realm of perceptual decisions. Taking advantage of the direct link between an optimal decision-making and the confidence in that decision, we offer a new dual-decisions method of inferring such confidence without asking for its explicit valuation. Our method circumvents the well-known miscalibration issue with explicit confidence reports as well as the specification of the cost-function required by opt-out or post-decision wagering methods. We show that observers inferred confidence in their first decision and its use in a subsequent decision (conditioned upon the correctness of the first) fall short of both the ideal Bayesian strategy, as well as of an under-sampling approximation or a modified Bayesian strategy augmented with an additional bias term to accommodate global miscalibration of confidence. The observed data are instead significantly better fitted by a model positing that observers use only few confidence levels or states, at odds with the continuous confidence function of stimulus level prescribed by a normative behavior. These findings question the validity of normative-Bayesian accounts of subjective confidence and metaperceptual judgments.