Modeling adaptive empathy based on neutral assessment: a way to enhance the prosocial behaviors of socialized agents under the premise of self-security

摘要

Ethical concerns over artificial intelligence (AI) have recently drawn extensive interest in both academia and industry. According to behavioral economics and neuropsychology, empathy may be an inherent mechanism to elicit prosocial behaviors. Therefore, we establish a three-layer general framework of empathetic AI (eAI) with emotion, empathy, and decision-making. By introducing different sub-models, three learning structures based on eAI are proposed, including a gradient ascent (GA)-based structure, an adaptive learning structure, and a practical learning structure. The dynamics of the first two structures are analyzed theoretically, and the practical dynamics are tested in games. We prove that in the prisoner’s dilemma (PD) environment, the GA-based eAI with neutral assessment can carry out adaptive cooperation and competition under the premise of self-security. In addition, although the modeling of empathy by extracting the emotional contagion and limited cognitive regulation is simplified and primitive, tests in the prisoner’s dilemma, the ultimatum game, and a multi-agent dilemma game, show that the eAI structure successfully elicits prosocial behaviors including altruism, cooperation and fairness. Compared with other socialized algorithms, the eAI structure has a more comprehensive coverage in terms of convergence, fairness, security, adaptability, and structural expansibility. Therefore, we believe this work can provide novel methods and insights for regulating the behaviors of socialized agents, as well as artificial subjects in psychological and economic experiments.