Does AI Crowd Out Context? Evidence from Clinical Radiology

Human–AI collaboration is often deployed on the premise that algorithms provide precision while human decision makers provide contextual judgment. The authors identify a limit to this complementarity. AI advice can reduce the value that humans derive from contextual information the algorithm does not observe. The authors study this in an experiment in which radiologists evaluate chest X-ray cases with varied access to AI predictions and clinical history. Without AI, clinical history improves agreement with a fully informed benchmark most in cases where its measured value is greatest. When AI is displayed with clinical history, this relationship reverses, and the combined workflow moves farther from the benchmark precisely where clinical history should contribute most. A Bayesian signal-extraction framework shows that some reduction in the contribution of clinical history is rational when AI increases the precision of the image-based information. Yet the observed reduction of clinical-history value is roughly 2.6 times the calibrated Bayesian precision-reweighting benchmark. The loss appears across diagnostic ambiguity levels and all but the most skilled radiologist groups, suggesting that simple targeting rules are unlikely to eliminate it. Behavioral evidence suggests that AI changes attention rather than effort. Radiologists remain active but engage less with clinical history, and their assessments stay closer to the AI prediction. The findings show that the value of human contextual knowledge is endogenous to workflow design. AI advice can crowd out human contextual knowledge rather than simply complement it. Effective AI deployment depends not only on model accuracy, but also on when and how AI advice enters the decision process.