Emotions Versus Laws as the Keys to the Ethical Design of Intelligent Machines

摘要

Whereas humans can know only about 200 other people well, machines will be able to know billions. This will define their higher level of consciousness and the threat they pose to humans. Trying to constrain the behavior of such machines by laws is like trying to define their behavior by a fixed set of expert system rules: it won't work. Rather, our focus should be the emotional values used to reinforce their learning of behavior. Their behaviors should be positively reinforced by happy humans and negatively reinforced by unhappy humans. Consciousness is a simulator for solving the temporal credit assignment problem in reinforcement learning, in the sense that consciousness enables brains to process experiences that are not actually occurring, Where time intervals between behaviors and rewards may be long and unpredictable, time intervals between simulations of those events can be short and predictable and thus amenable to the brain's known mechanisms for solving the temporal credit assignment problem. The higher level consciousness of machines will simulate human social interactions in detail, in order to learn complex behaviors for coping with conflicts among humans and other ambiguities of human happiness.