Overcoming medical overuse with AI assistance: An experimental investigation

This study examines the role of Artificial Intelligence (AI) in reducing medical overtreatment, a critical healthcare challenge that increases costs and patient risks. In two experiments - with 196 physicians at a hospital and 120 students at a medical school in Wuhan - we use a novel medical prescription task under three incentive schemes: flat (constant pay), progressive (pay increases with treatment quantity), and regressive (penalties for overtreatment) to estimate receptivity to AI assistance and its effects on overtreatment and treatment accuracy, and test whether effects vary with incentives. AI recommendation of a treatment is estimated to increase the probability a physician prescribes it by 25.7-28.4 percentage points (pp), with the largest effect under the flat scheme. Physicians are more receptive to AI recommendations in medical domains with which they are less familiar. We estimate that AI assistance reduces the probability a physician overtreats by 10.9-25.7 pp (15.2-80.3%), with significantly larger absolute and relative effects under the flat scheme compared to progressive and regressive schemes. AI assistance improves physicians' treatment accuracy by 9.8-13.3 pp (14.6-19.9%), with the largest absolute effect under the regressive scheme. These findings are corroborated by the medical school experiment, which reveals that factors indicative of insufficient ability account for 34% of the explained variation in overtreatment, monetary incentives account for 22%, patient welfare considerations account for 20%, and factors related to defensive medicine for 10%. These results provide valuable insights for healthcare administrators considering AI integration into healthcare systems.