Transparency features alone are insufficient for trust calibration in artificial intelligence diagnostic support.

Abstract

Objective: Clinician trust in artificial intelligence (AI) is a frequently cited barrier to adoption, yet there is limited empirical evidence on how transparency features influence trust. This study aimed to address this evidence gap.

Materials and methods: We conducted a randomized controlled experiment in which 220 clinicians were randomized to 1 of 3 conditions: no transparency (control), numeric confidence scores, or narrative explanations with citations, using simulated AI diagnostic assistance across clinical case vignettes.

Results: Transparency did not influence tool-specific trust [F(2, 214) = 0.38, P = .68], global trust [F(2, 217) = 0.13, P = .88], or perceived information credibility [F(2, 210) = 0.42, P = .66]. Confidence increased following any AI assistance (+6.7%), but narrative explanations were associated with greater change in confidence compared with both control (P = .024) and numeric confidence (P = .0051). Participants changed or refined diagnoses after reviewing AI output 45.8% (SD 23.4%) of the time, with participants in the control group demonstrating higher levels of behavioral influence than transparency conditions (P = .0006).

Discussion: Trust, confidence, and behavioral influence represent distinct constructs and are impacted differently by transparency methods. The difference in these contributions to clinician-AI partnerships should be considered when designing for safe trust calibration, with particular consideration for the dissociation between trust and confidence.

Conclusion: These findings suggest common interface transparency features alone failed to produce a detectable, meaningful shift in clinician trust in AI assistance and that changes may instead alter behavioral influence and confidence without a corresponding change in trust.