ELIZA fifty years later: An automatic therapist using bottom-up and top-down approaches

Abstract

Our methods for realizing a moral artificial agent assume that the wisdom of crowds can equip a machine with the enormous number of experiences that are the source of its ethical reasoning. Every second, people with different cultural, religious or social backgrounds share their personal experiences about multitudes of human acts. We propose that a machine therapist capable of analyzing thousands of such cases should be more convincing and effective talking to a patient, instead of analyzing single keywords. In this chapter, we introduce this vision and several techniques already implemented in an algorithm for generating empathic machine reactions based on emotional and social consequences. We show the roles that Bentham’s Felicific Calculus, Kohlberg’s Theory of Stages of Moral Development and McDougall’s classification of instincts play in the agent’s knowledge acquisition, and we describe the accuracy of already working parts. Modules and lexicons of phrases based on these theories enable a medical machine to gather information on how patients typically feel when certain events happen, and what could happen before and after actions. Such empathy is important for understanding the actions of other people, and for learning new skills by imitation. We also discuss why this bottom-up approach should be accompanied by a top-down utility calculation to ensure the best outcome for a particular user, and what ethical dilemmas an advanced artificial therapist could cause.