Preventing Sociopathic Robots

Preventing Sociopathic Robots is an AI safety research program that explores how to design world-modeling AI agents whose prosocial alignment emerges from their own homeostatic needs rather than from external constraints alone. Using simulations in an open-ended, Minecraft-like environment, the team studies agents endowed with meta-learning, counterfactual modeling, and explicit representations of vulnerability and bodily-like needs to see when cooperative or antisocial “personalities” develop. The project builds on prior theoretical work arguing that artificial empathy must incorporate affect and embodied vulnerability and has produced a Science Robotics paper proposing architectural principles for preventing antisocial machine behavior.