Ava Thomas Wright's website Thu, 02/08/2024 - 4:00pm 115 Peabody Abstract for "Humanity Compatible AI": Computer scientist Stuart Russell criticizes the “standard model” in which artificially intelligent (AI) agents are designed to learn how to optimize the achievement of objectives that we specify for them. The problem with this model, Russell argues, is that AI agents may pursue those objectives in unforeseen ways that violate our values and outstrip our control. Russell argues that AI agents should be designed, instead, to learn what our objectives are by inferring them from its observations of what we do or say. AI agents designed in this way would thus continuously align their behavior with our values and defer to our control, Russell argues. While I believe that Russell’s new model for “human compatible” AI is an important innovation, the model raises a new central question, "How should we program AI agents to interpret human behavior?" Russell favors an economic approach that interprets our behavior as efforts to maximize a utility function composed by our rationalized preference rankings of all possible future lives. In this paper, I criticize this approach and argue that AI agents should determine our objectives by interpreting our behavior as efforts to act autonomously, instead. Immanuel Kant refers to our capacity for free, autonomous choice as our “humanity,” and it is the foundation of our moral and legal rights. I thus argue for humanity compatible AI. Ava Thomas Wright, JD, MS, PhD is an assistant professor of philosophy and affiliated faculty in computer science at Cal Poly San Luis Obispo, specializing in AI ethics, ethics, and legal and political philosophy. A former technology business founder with an MS in artificial intelligence as well as a JD and a PhD in philosophy, Wright’s expertise is broad and deep in ethics, political philosophy, law, business, and AI. Wright presents frequently on AI ethics topics and has published on AI ethics and law in the anthologies Technology Ethics (Routledge) and Kant and Artificial Intelligence (De Gruyter). Wright’s current research in AI ethics focuses on aligning the behavior of highly autonomous AI agents with Kantian principles of ethics and justice.
