Events

Prof. Ricardo Alonso
Professor of Management
London School of Economics and Political Science

We study how a privately informed decision maker uses recommendations from an AI system. We treat the AI as a black box guaranteeing a minimum level of performance when its advice is followed. A conservative Bayesian decision maker combines the AI recommendation with her signal to maximize the worst-case performance consistent with that guarantee. By mapping this robust inference problem into a Bayesian persuasion problem, we characterize the conditions under which the decision maker ignores the AI, combines its recommendation with her own information (“complementary AI”), or fully delegates to it (“substitutive AI”). By reinterpreting the robust inference problem as an optimal reliance problem, we derive explicitly the value of the AI’s recommendation. We show that the value of AI depends on the form of expertise. For example, in a classification problem, a prescriptive expert—one who knows which class is more likely—typically either ignores the AI or delegates to it, whereas a proscriptive expert—one who knows which class is less likely—always finds AI complementary and benefits more from its advice. We characterize which types of experts benefit the most from Human-AI interaction.