Perceptual confidence neglects decision-incongruent evidence in the brain

Our perceptual experiences are accompanied by a subjective sense of certainty. These confidence judgements typically correlate meaningfully with the probability that the relevant decision is correct 1 , 2 , 3 , 4 , 5 , 6 , bolstering prevailing opinion that both perceptual decisions and confidence optimally reflect the probability of having made a correct decision 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 . However, recent behavioural reports suggest that confidence computations overemphasize information supporting a decision, while selectively down-weighting evidence for other possible choices 14 , 15 , 16 , 17 , 18 , 19 . This view remains controversial, and supporting neurobiological evidence has been lacking. Here we use intracranial electrophysiological recordings in humans together with machine-learning techniques to demonstrate that perceptual decisions and confidence rely on spatiotemporally separable neural representations in a face/house discrimination task. We then use normative computational models to show that confidence relies excessively on evidence supporting a decision (for example, face evidence for a ‘face’ decision), even while decisions themselves reflect the optimal balance of all evidence (for example, both face and house evidence). Thus, confidence may not reflect a readout of the probability of being correct; instead, observers may sacrifice optimality in favour of self-consistency 20 in the face of limited neural and computational resources. Although seemingly suboptimal, this strategy may reflect the inference problem that perceptual systems are evolutionarily optimized to solve.