0.6 Social interaction as a discovered world

The social world is not a single dynamical system observed from outside. It is populated by other learners with private information, changing goals, and their own models of the interaction. Gaze direction, turn-taking, imitation, joint attention, refusal, and repair are fragments from which the infant must infer what kinds of agents and relationships are present.

Spelke’s distinction between goal-directed agents and social beings is important here. Efficient motion toward an object may support a prediction of goal-directed action; reciprocal attention or affect supports different expectations about social engagement. Neither system gives the infant a complete theory of a particular person. They provide typed probes through which repeated interaction can be organized.

Machine theory-of-mind systems illustrate one computational analogue: a model observes agents’ behavior and learns to predict their future actions from their inferred characteristics and mental states [ Rabinowitz et al. , 2018 ] . Prediction is valuable, but it is not yet causal, normative, or trustworthy interpretation. An ORACLE learner must also preserve the provenance of an inference: which observations support the attribution, which other agents supplied information, and which conclusions must be repaired when an expectation fails.

Social accommodation is especially delicate. If another person acts unexpectedly, the learner might revise a belief about that person’s goal, a general rule about social conduct, or its own interpretation of the episode. These repairs have different consequences. A viable theory must localize the defect rather than globally overwriting everything learned about the person or the practice.