Observing others stay or switch – How social prediction errors are integrated into reward reversal learning

Reward properties of stimuli can undergo sudden changes, and the detection of these 'reversals' is often made difficult by the probabilistic nature of rewards/punishments. Here we tested whether and how humans use social information (someone else's choices) to overcome uncertainty during reversal learning. We show a substantial social influence during reversal learning, which was modulated by the type of observed behavior. Participants frequently followed observed conservative choices (no switches after punishment) made by the (fictitious) other player but ignored impulsive choices (switches), even though the experiment was set up so that both types of response behavior would be similarly beneficial/detrimental (Study 1). Computational modeling showed that participants integrated the observed choices as a 'social prediction error' instead of ignoring or blindly following the other player. Modeling also confirmed higher learning rates for 'conservative' versus 'impulsive' social prediction errors. Importantly, this 'conservative bias' was boosted by interpersonal similarity, which in conjunction with the lack of effects observed in a non-social control experiment (Study 2) confirmed its social nature. A third study suggested that relative weighting of observed impulsive responses increased with increased volatility (frequency of reversals). Finally, simulations showed that in the present paradigm integrating social and reward information was not necessarily more adaptive to maximize earnings than learning from reward alone. Moreover, integrating social information increased accuracy only when conservative and impulsive choices were weighted similarly during learning. These findings suggest that to guide decisions in choice contexts that involve reward reversals humans utilize social cues conforming with their preconceptions more strongly than cues conflicting with them, especially when the other is similar.