Subjects took part in separate behavioural and fMRI sessions on the same day. The prescan session involved four phases. In phase 1, we calibrated four levels of coherence so as to achieve target levels of choice accuracy (60%, 70%, 80% and 90%). In phases 2–4, we trained subjects on the social task. In phase 2, subjects were paired with the partners in a block-wise manner (each partner is indicated by a unique colour and name). There were four cycles of blocks of 10 trials per partner (e.g., A–B–C–D–A–B–C–D–A–B–C–D–A–B–C–D). The context screen was shown before each block of trials but not after a perceptual decision. In phase 3, subjects were paired with the four partners in an interleaved manner, with the current partner’s identity revealed after each perceptual decision. In phase 4, the ‘showdown’ stage was played out in the background. In addition, we introduced a condition where the social context was hidden. The scan session involved four runs, using the same design as in phase 4 of the prescan session. We matched the distribution of conditions (coherence × context) across scan runs in order to facilitate multivariate analysis of the fMRI data.