Chenhao Tan (UChicago) Human-centered AI: How to Generate Useful Explanations for Human-AI Decision Making

Abstract: Human-centered AI advocates the shift from emulating human to empowering people so that AI can benefit humanity. A useful metaphor is to consider human as a puzzle piece; it is important to know the shape of this puzzle piece so that we can build AI as a complement. In this talk, I focus on the case of AI-assisted decision making by offering explanations of predictions to illustrate key principles towards human-centered AI. Ideally, explanations of AI predictions enhance human decisions by improving the transparency of AI models, but my work reveals that current approaches fall short of this goal. I then develop a theoretical framework to show that the missing link lies in the neglect of human interpretation. I thus build algorithms to align AI explanations with human intuitions and demonstrate substantial improvements in human performance. To conclude, I will compare my perspective with reinforcement learning from human feedback and discuss further directions towards human-centered AI.