When AI Disagrees: The Effect of Second Opinion on Patients’ Trust in Doctors

Chen, C., Sun, Y., Mengqi (Maggie) Liao, & Sundar, S. S. (2026). “When AI Disagrees: The Effect of Second Opinion on Patients’ Trust in Doctors.” International Journal of Human-Computer Studies, 103824. https://doi.org/10.1016/j.ijhcs.2026.103824  Abstract: As we increasingly turn to AI tools for second opinions, how might that affect our trust in doctors? We study this in the context of mental health consultation and medical advice, by examining how agreement or disagreement by an AI system influences patients’ trust in their doctors. To answer this question, we conducted an experiment (N = 135) in which participants interacted with a Large Language Model (LLM)-simulated doctor during a stress management therapy session. At the end of the consultation, the doctor offered to consult an AI assistant for a second opinion. Results showed that AI agreement had mixed effects: it increased perceptions of doctor laziness while also enhancing perceived recommendation credibility. In contrast, AI disagreement increased perceived medical uncertainty and subsequently reduced patients’ trust. Since patients may attribute varying levels of anthropomorphism to an LLM-simulated doctor, when the doctor was perceived as more human-like, AI disagreement increased perceived medical uncertainty, whereas AI agreement increased both perceived doctor laziness and perceived recommendation credibility. These changes in perception were associated with cognitive, affective, and behavioral trust in the doctor. We discuss these findings to advance theoretical understanding of agency negotiation within doctor-patient-AI interactions, offer practical suggestions for how doctors can better communicate AI disagreement and agreement to patients, and provide methodological insights for leveraging LLMs for experimental design and message manipulation. Limitations of using LLMs to impersonate human professionals, as well as the limited generalizability of the findings derived from LLM-generated role-play, are also discussed.