Preventing Sociopathic Robots

We will run a variety of simulations to demonstrate how prosocially aligned AI may be emergently bootstrapped via vulnerable engagement (alone and in competition/collaboration with others) with an open-ended Minecraft-like virtual world.

Such environments are well-suited for training and testing for our systems, who have several planned emergent capacities: (I) Abilities for meta-learning and counterfactual modeling; (II) Vulnerability; (III) Incorporation of biologically-inspired inductive biases.

Combined with meta-learning and counterfactual modeling that can accommodate complex agent-based relationships across multiple time scales, we will demonstrate how social alignment may emerge as a kind of predictive homeostasis (i.e., allostasis, or “sociostasis”).

In close collaboration with other groups dedicated to human-centered AI, we will focus on determining the extent to which human-mimetic design may allow for the development of effective and robustly beneficial artificial intelligences.

We hope to a) determine conditions that are more or less likely to result in either prosocial or antisocial ‘personalities’, b) evaluate the consequences of these agent traits on interpersonal interactions, and c) explore the degree to which these emergent social dynamics alter the personalities of our agents.