Generative Artificial Intelligence (AI)–Assisted Self-Diagnosis and the Patient-Physician Relationship: Mixed Methods Study of Calibration, Participation, and Trust

Abstract

Abstract Background Patients increasingly use generative AI to interpret symptoms and seek health information, yet limited evidence shows how AI-assisted self-diagnosis is integrated into care-seeking and related to clinical interactions and patient–physician relationships. Objective This study examined how AI-assisted self-diagnosis is incorporated into care-seeking processes and how it relates to patient participation in clinical encounters and trust in physicians. Methods An exploratory sequential mixed methods design was used. In the qualitative phase, Chinese adults who had used generative AI to interpret symptoms, appraise possible conditions, or seek health advice within the previous year were purposively recruited through Xiaohongshu, WeChat Moments, and WeChat groups. Semistructured interviews were conducted from August 20, 2025, to January 10, 2026, and analyzed using reflexive thematic analysis. In the quantitative phase, an anonymous web-based survey was conducted in China through Huixiang Data from January 25, 2026, through January 28, 2026. Adults who had used generative AI for health consultation involving symptom interpretation or preliminary self-diagnosis within the previous 6 months were recruited through convenience sampling. Measures included perceived AI-assisted self-diagnosis quality, calibrated illness appraisal, patient participation, diagnosis validation, diagnosis comprehension, trust in physicians, AI use frequency, trust in health information sources, and demographic characteristics. Trust in physicians was assessed using 5 adapted items covering competence, integrity, and benevolence. Descriptive statistics, Pearson correlations, and PROCESS mediation analyses were performed. Results Qualitative findings (n=48) indicated that AI-assisted self-diagnosis was commonly used in a prediagnostic gray zone for preliminary orientation, informal triage, and interim self-management. Participants described AI as helping them appraise illness severity, prepare for consultations, ask questions, and understand physicians’ diagnoses and reasoning. Quantitative findings (n=546) were consistent with these patterns. Perceived AI quality was positively associated with calibrated illness appraisal (b=0.57, 95% CI 0.49-0.64), which was positively associated with patient participation (b=0.35, 95% CI 0.28-0.43). The indirect association was significant (estimate=0.20, 95% bootstrap CI 0.14-0.26). Perceived AI quality was also associated with diagnosis validation (b=0.69, 95% CI 0.61-0.76) and diagnosis comprehension (b=0.65, 95% CI 0.58-0.73), which were associated with trust in physicians (b=0.15, 95% CI 0.07-0.23 and b=0.20, 95% CI 0.12-0.28, respectively). The corresponding indirect associations were 0.10 (95% bootstrap CI 0.04-0.17) through diagnosis validation and 0.13 (95% bootstrap CI 0.07-0.19) through diagnosis comprehension. Conclusions Generative AI may function as an informational intermediary across the care-seeking process rather than undermine medical authority. By supporting illness appraisal, consultation preparation, and postconsultation understanding, AI-assisted self-diagnosis may be associated with greater patient participation and trust in physicians.