Emotionally-Sensitive AI-driven Android Interactions Improve Social Welfare Through Helping People Access Self-Transcendent States

Social humanoid robots are generally built with the purpose of helping individuals achieve difficult tasks or feel less lonely. But a novel use of humanoid robots, especially robots drawing on emotionally sensitive AI, is to use the context of what feels like a human relationship to help people practice reaching advanced stages in human development. At the peak of Maslow’s hierarchy of needs and models of self-development is the state of self-transcendence, which includes expansive feelings of love. Although beings can have difficulty reaching this state, several lines of research have shown that even briefly accessing states of self-transcendence can improve physical and psychological well-being. In this paper we briefly present results of the first experiments of which we are aware in which AI-driven, audiovisual, interactive android technology is successfully used to support the experience of self-transcendence. Individuals had AI-driven conversations with emotionally responsive AI embedded in a humanoid robot, its audio-visual avatar, or audioalone avatar. These conversations were based on exercises reported to induce self-transcendence in humans. In experiment 1, we tested an initial version of this AI using brief, constrained interactions with Sophia the humanoid robot and no emotion detection (N=26). In experiment 2, we tested a more sophisticated version of this AI including deep-learning-based emotion detection deployed in the context of a slightly longer and slightly less constrained interaction. Conversations were with either Sophia or one of two avatars (one with a face and voice, the other with only a voice; N=35). The results suggest that conversations between humans and a humanoid robot or its audiovisual avatar, controlled by emotionally responsive AI, are accompanied by self-transcendent emotions. Most importantly, objective correlates of those emotions are detectable by a deep learning network. 32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada.