When Doctors and AI Disagree: The Effect of the Doctor’s Explanation Strategy on Patient Responses

Large Language Models (LLMs) are increasingly shaping how patients prepare for medical consultations. Patients now enter consultations with health advice generated by Artificial Intelligence (AI), creating a triadic doctor-patient-AI relationship in which a medical professional’s clinical expertise competes with AI as an epistemic source. Managing situations involving conflicting diagnoses and subsequent treatment advice generated by AI poses a significant challenge for medical professionals nowadays. This thesis addresses the outlined problem by examining how doctors should explain their diagnosis when AI-generated diagnosis and treatment advice provided by patients diverge from their clinical judgment. Drawing on Lay Epistemic Theory, this research argues that counterfactual explanations, which explain why the AI-generated alternative does not fit the patient’s presentation of symptoms, are more effective than mechanistic explanations, which explain what causes the observed symptoms. Across three between-subjects vignette experiments, this research tests the effect of the doctor’s explanation strategy on patients’ evaluation of the medical encounter and their behavioural intentions. Study 1 shows that a counterfactual explanation increases patients’ consultation satisfaction compared with a mechanistic explanation. Study 2 replicates this effect of the doctor’s explanation strategy on patient satisfaction, extends it to trust in the doctor as an alternative dependent variable, and shows that perceived epistemic authority mediates these effects. Study 3 examines the length of the doctor-patient relationship as a boundary condition and extends the dependent variables to include behavioural intentions. A counterfactual explanation strengthens perceived epistemic authority, especially when the doctor-patient relationship is short, which in turn increases consultation satisfaction, trust in the doctor, and the willingness to follow the doctor’s recommendation, while reducing the willingness to follow the AI’s recommendation and the intention to reuse the AI. The findings of this thesis contribute to research on medical AI, health communication, Lay Epistemic Theory, and credence services by showing that doctors can reconstruct epistemic authority through contrastive explanation in AI-conflicted consultations.