Fig. 1: Methods and behaviour.

a Trials consisted of three time-windows: the decision-window; bet-window; and feedback-window. In the decision-widow, a fixation cross was presented for a variable duration, followed by 350 ms of random dot motion, and up to 1 s to enter the left/right response, with visual feedback about the button pressed. In the bet-window, participants were given up to a second to bet that their response was correct, to double the value of the points gained/lost on that trial. In the feedback-window, participants were cued (coloured fixation) about whether they would receive explicit feedback, or be shown two question marks indicating they should think about how many points they think they won/lost on that trial. If participants failed to enter their left/right decision within one second after the stimulus, they were informed with the words ‘time out’, they lost a point, and the trial was excluded from the analysis. b Coherence (percent of dots moving in the correct direction) was adjusted each block (50 trials) to maintain accuracy around 75% correct. Line shows the average, shaded region shows 95% within-subject confidence intervals. c Average sensitivity (d’) across blocks, with 95% within-subject confidence intervals shaded. The red dashed line shows the fit of the best fitting model, which uses confidence to learn on every trial. The blue dashed line shows how sensitivity would decrease with decreased stimulus coherence simulating observers who do not learn. d Sensitivity (d’) on bet (ordinate) and no-bet (abscissa) trials. e Median reaction time (s) on correct trials for bet and no-bet trials. f Perseveration, an average of the normalised probability of repeating a response, for repeat (filled) and alternate (open) stimuli, by feedback on the previous trial, or the previous bet response on no-feedback trials (error bars show 95% CI within-subjects difference between repeat and alternate stimuli). The markers show the predictions of the computational model using confidence to learn. All panels include N = 23 participants, source data are provided as a Source data file.