Abstract
The Bayesian confidence hypothesis, which postulates that confidence reflects the posterior probability that a decision is correct, is currently the most prominent theory of confidence. Although several recent studies have found evidence against a strictly Bayesian account in the context of relatively complex tasks, the Bayesian confidence hypothesis remains dominant for simpler tasks. However, confidence could instead reflect a simpler distance-to-criterion computation where confidence is determined by the distance between the sensory evidence and the decision criterion. Here, we uncover a basic behavioral signature that distinguishes Bayesian from distance-to-criterion models for simple 2-choice tasks. Specifically, if confidence reflects the probability of being correct, then confidence criteria in conditions of varying difficulty should have constant log odds but differ in evidence space. Alternatively, if confidence follows a distance-to-criterion computation, then confidence criteria in conditions of varying difficulty should be fixed in evidence space but have different log odds. We examined this signature in three different datasets, including two that have previously been used to support the Bayesian confidence hypothesis. All three datasets exhibited behavioral signatures in line with distance-to-criterion but contrary to Bayesian computations. We further performed extensive comparison of 32 models that implemented either Bayesian or distance-to-criterion confidence computations and systematically differed in their auxiliary assumptions. These model comparisons provided overwhelming support for the distance-to-criterion models over their Bayesian counterparts across all model variants and across all three datasets. Specifically, distance-to-criterion models had average AIC values 110.2, 68.1, and 70.2 points lower than Bayesian models across the three experiments, respectively. Finally, we traced the discrepancy with previous results supporting the Bayesian model to the details of the fitting procedure and also uncovered a previously unappreciated mimicry between metacognitive noise and lapse rate parameters. These observations falsify the Bayesian confidence hypothesis and bring important insights into many components of confidence computations.