Reverse Engineering the Intelligence of Care
-----
Many animals care for their young, but the human caregiving system is distinct and adaptive. Caregiving roles, intensity, and institutions vary across cultures and time. Still, caregivers enable helpless infants to survive, learn, and ultimately develop into autonomous individuals. What computational principles make human caregiving both so flexible and so successful? How might we build the machines that will someday care for us? Kleiman-Weiner will present a computational cognitive model of caregiving formalized in the language of sequential decision-making. In this model, the caregiver chooses actions that maximize the expected utility of a learner at a future point when the caregiver is no longer present. The model distinguishes care from helping and teaching: caregivers enable safe exploration that builds a learner's capacity and autonomy over time. In experiments with parent-child dyads, the model quantitatively explains when caregivers decide to step in to help and when they hold back to let the child persevere. Finally, researchers find that today's AI agents intervene earlier and more often than people do, risking disempowerment rather than building autonomy.
-----