Events

Prof. Kevin Bryan
Associate Professor of Strategic Management
Rotman School of Management
University of Toronto

AI predicts; its prediction is used by humans in decisions. The value of AI therefore depends on how humans extend, verify, and act on that prediction. We model AI as part of a composite experiment where agents can verify predictions at cost, delegate to AI, or avoid the model altogether. We derive the optimal coverage–conditional accuracy tradeoff for training, showing that maximizing unconditional accuracy is generally suboptimal. The optimum is discontinuous where users switch between autonomous and verified AI regimes. Heterogeneous users disagree on the ideal model. Because the optimal model maps directly from the economic environment – downside risk, verification cost, adversarial pressure, task complementarities – knowledge of the economic environment can inform about the nature of optimal training ex-ante.