This chapter contributes to the ongoing debate by considering how best to interpret Prediction Error Minimization (PEM) models in the case of social cognition and in particular mindreading. It suggests that PEM is promising as a general computational account in Marr's sense, that is as an abstract account of the formal properties of the psychological mechanisms that carry out social cognition and in particular mindreading. The chapter recapitulates the trajectory of mindreading research in recent decades, and considers some recent developments that have grown out of the frustration with earlier theories. It asks how theories and models in social cognition research may be evaluated along three dimensions corresponding to the three levels of analysis proposed by Marr: computational, algorithmic, and implementational. The chapter considers how a PEM framework may be used to provide structure and direction for research on mindreading in particular and possibly of social cognition generally.