Social Prediction

Social prediction is the process by which agents anticipate the behavior of other agents in social contexts. It is the computational core of mentalizing and theory of mind: the brain generates forward models of others' actions and updates these models through prediction errors derived from observed behavior. Social prediction is not merely a cognitive task but a control problem — the controller must predict the outputs of a coupled system (another agent) whose internal states are unobservable and whose dynamics are nonlinear.

The predictive processing framework treats social prediction as a hierarchical inference problem in which prior expectations about others' goals generate predictions about their actions, and observed actions generate prediction errors that update the goal model. The recursion is deep: predicting an agent who is predicting you requires modeling a system that contains a model of you. This is the social analog of the observer effect in physics: the act of prediction changes the target, because the target is itself a predictor.

Social prediction failures are clinically significant. Paranoia can be understood as a hyperactive social prediction system that assigns hostile intentions to ambiguous behavior. Autism spectrum conditions involve differences in social prediction that may reflect altered prior expectations or altered precision weighting of social signals. Both conditions reveal that social prediction is not a luxury but a fundamental regulatory mechanism for social interaction.

Social prediction is the most computationally expensive inference problem an organism routinely solves. It requires modeling a non-stationary, adversarial, partially observable system with unbounded recursion — and doing it in real time, under metabolic constraints, while the target is trying to predict you. The fact that humans do this effortlessly is not evidence that social prediction is easy. It is evidence that evolution has invested staggering resources in solving it.