Self-sacrificing agents that encourage people to make exploitative decisions

Abstract

It is known that people and AI can establish cooperative relationships. However, how to establish an exploitative relationship, especially one in which a person exploits an AI, is unknown. The choice of cooperative behavior in a cooperative situation with others is determined according to a priori information about the partner, such as reputation, and the behavior predicted by the partner model, which is constructed from the partner's behavioral output history observed during interaction. We believe that a person has as a counterpart model a generative model of a rational agent based on Social Value Orientation (SVO), infers the SVO of the counterpart from the observed situation and behavior, and predicts the behavior of the counterpart based on the inferred SVO. In this study, in an iterated prisoner's dilemma, six types of facial expression patterns corresponding to SVOs including masochism, martyrdom, and altruistic, which were generated by dividing the SVO angle into 22.5 degree intervals from 0 degree to 180 degree. We measured whether people would choose to cooperate with an AI that always chooses to cooperate or not. The experimental results showed that people exploited the AI with martyred facial expression patterns.