Conference on Cognitive Computational Neuroscience

A "positive evidence bias" (PEB) is the empirical observation that confidence ratings (unlike first-order decisions) are relatively insensitive to evidence against a selected perceptual choice. Such findings have been interpreted as human metacognition being limited by biases or heuristics. However more recent work has shown that a PEB emerges for high-dimensional stimulus sets when unitary neural network architectures are trained to track the correctness of their decisions (Webb et al., 2022). Here we show that Bayes-optimal confidence also shows a PEB in higher-dimensional signal detection theoretic spaces. This effect occurs due to a nonlinearity induced by normalisation of confidence by a large number of unchosen alternatives, and also emerges in neural network models when the number of unchosen options in a stimulus set. Our analysis suggests that a PEB is rational if subjects are computing confidence in an evidence space that is larger than that assumed by the experimenter.