Artificial intelligence (AI) doctors are chatbots that diagnose patients on the Internet using algorithms to determine patients’ mental condition and need for medical treatment. I build a mathematical model to investigate how AI doctors affect patients and the efficiency and utilization of healthcare. Intuition suggests that AI doctors improve patients’ healthcare utilization decisions and improve healthcare efficiency and patients’ well-being. My analysis reveals that these intuitions only hold when AI doctors’ diagnoses are very accurate. When AI doctors are not very accurate, their impacts on patients and healthcare vary with patient types (the severity of their mental condition and the cost they face to visit hospitals). Specifically, AI doctors change patients’ decisions to go to hospitals when patients are moderately uncertain about their need to go there. AI doctors increase healthcare utilization by patients with low access and discourage utilization by patients with easy access. AI doctors benefit patients with moderate access to hospitals, severe patients with low access, and mild patients with easy access to hospitals; however, AI doctors harm mild patients with low access and severe patients with easy access to hospitals. Moreover, patients are better off knowing AI doctors' true accuracy; neither overtrust nor undertrust in AI doctors improves healthcare. These findings caution hospitals and public health policymakers that adopting AI for mental health diagnosis can backfire to harm patients and exacerbate healthcare inefficiency.  

Keywords: artificial intelligence, healthcare, public health, patients, diagnosis