In this seminar, Yue Zhang will give a talk about AI-Augmented Human Decision Making: Belief Formation, Choice, and Evidence from Medical Diagnosis.
Human decision makers (HDMs) increasingly rely on artificial intelligence (AI) systems for decision support in high-stakes contexts, yet how HDMs integrate AI recommendations into their decision-making process remains insufficiently understood. Observed responses to AI assistance confound how humans form beliefs from available information with how they translate those beliefs into decisions, making it difficult to identify the behavioral mechanisms underlying human-AI collaboration. We develop a structural model of AI-augmented human decision making that separates belief formation from stochastic choice. The framework combines a general belief-aggregation layer, which allows human and AI evidence to receive different weights and interact, with a stochastic choice layer that captures imperfect translation of beliefs into actions. The analysis shows that belief distortion and choice noise represent distinct but interacting behavioral mechanisms: while deterministic decisions can be insensitive to small belief distortions, stochastic choice makes decision probabilities continuously responsive to belief distortions. We estimate the model using experimental data from a medical diagnostic task in which endoscopists classify colonoscopy cases with and without AI assistance. The results show that HDMs incorporate both their own diagnostic evidence and AI recommendations, but systematically underweight AI recommendations relative to the normative Bayesian benchmark. Moreover, compared with non-experts, expert HDMs place greater weight on their own diagnostic evidence while maintaining similar responsiveness to AI, indicating that expertise affects the belief-aggregation process itself rather than merely the strength of human evidence. Finally, by recovering the full choice probability rather than only realized decisions, the model enables evaluation of AI’s decision impact beyond observed sample paths and a probabilistic assessment of AI-assisted diagnostic performance.