Schematic overview of the experimental design(s). (a) Each study had a first fMRI session dedicated to acquire the data necessary to construct a ‘decoder’, a machine learning classifier of brain activity patterns. (b) All studies shared the same basic trial design: a first period of 6 seconds called ‘induction’, during which participants were instructed to “modulate, or change their brain activity in order to maximize the feedback received later in the trial”. Unbeknownst to participants, brain activity patterns were fed to the previously trained decoder to calculate the likelihood that the current brain state corresponded to the target state. The likelihood was fed back to participants in the form of a disc within a circle: the larger the disc, the larger the obtained reward. The circle indicated the maximum attainable reward. Beyond this basic structure, the design was slightly modified to meet the requirement of each individual study. In study 1, a neutral face was presented for 0.5 second before the onset of the induction period. Note that in the original study actual face images were used, not cartoons. In study 2 and 3, instead of a simple visual cue indicating the induction period, participants were presented with an achromatic grating. Lastly, in study 4, a random dot kinetogram with 0% coherence was shown for 2 seconds at variable onsets during the induction period.