Observers translate information about other agents' higher-order goals into expectations about their forthcoming...

Social perception relies on the ability to understand the higher-order goals that drive other people's behaviour. Under predictive coding views, this ability relies on a Bayesian-like hypothesis-testing mechanism, which translates prior higher-order information about another agent's goals into perceptual predictions of the actions with which these goals can be realised and tests these predictions against the actual behaviour. We tested this hypothesis in three preregistered experiments. Participants viewed an agent's hand next to two possible target objects (e.g., donut, hammer) and heard the agent state a higher-order goal, which could be fulfilled by one of the two objects (e.g., "I'm really hungry!"). The hand then reached towards the objects and disappeared at an unpredictable point mid-motion, and participants reported its last seen location. The results revealed the hypothesized integration of prior goals and observed hand trajectories. Reported hand disappearance points were predictively shifted towards the object with which the goal could be best realised. These biases were stronger when goal statements were explicitly processed (Experiment 1) than when passively heard (Experiment 2), more robust for more ambiguous reaches, and they could not be explained by attentional shifts towards the objects or participants' awareness of the experimental hypotheses. Moreover, similar biases were not elicited (Experiment 3) when the agent's statements referred to the same objects but did not specify them as action goals (e.g., "I'm really not hungry!"). These findings link action understanding to predictive/Bayesian mechanisms of social perception and Theory of Mind and provide the first evidence that prior knowledge about others' higher-level goals cascades to lower-level action expectations, which ultimately influence the visuospatial representation of others' behaviour.