Why Doctors Trust Transparent AI (And Reject the Black Box)

A 2025 study from the World Economic Forum tracked how clinicians responded to AI diagnostic suggestions across two conditions: a proprietary black-box system and an open-source model whose reasoning could be inspected. The results are stark. When the AI was opaque—outputting a prediction without showing its work—clinicians overrode its recommendations 73% of the time. When the same clinicians used a transparent model, they overrode it only 1.7% of the time. The usual reading of this data is that clinicians need more "AI literacy" or training to trust algorithmic tools. That interpretation assumes the doctors are the problem. But look at the asymmetry: the same professionals, the same clinical setting, the same task—the only variable was whether they could see how the machine arrived at its conclusion. A 1.7% override rate means they found the transparent AI's reasoning sound enough to act on in over 98% of cases. That is not a trust deficit. That is rational calibration. Here is the mechanism the industry prefers to ignore. When a black-box system flags a potential stroke risk, the clinician cannot distinguish between a true signal and a spurious correlation—say, a model that learned a hospital-specific pattern (e.g., "patients in room 312 tend to have worse outcomes") rather than a genuine clinical marker. The only rational response is to treat the prediction as noise until proven otherwise. The 73% override rate is not stubbornness; it is professional accountability. You cannot be responsible for a decision whose reasoning you cannot audit. The transparent model, by contrast, invites the clinician into the loop. They can see which features the model weighted, check whether those features are clinically plausible, and correct the model when it misfires on edge cases—a patient with atypical lab values, a rare comorbidity the training data under-sampled. This is not about making doctors feel better. It is about giving them the tools to perform the diagnostic reasoning they were trained for, augmented rather than overridden by computation. This shifts the regulatory question. The current policy battles in medical AI center on restricting openness: proposals to audit or gatekeep model weights, to limit what developers can disclose, to treat transparency as a security risk. But the WEF data suggests the opposite direction: the trust problem is a design problem. Transparency does not just make clinicians feel more comfortable; it doubles as a performance upgrade for the entire human-AI system. When doctors can see the reasoning, they catch errors the model would miss and apply judgment the model lacks. The override rate drops because the system works better together. The real challenge is not getting doctors to trust AI. It is building AI that deserves their trust by design, not by declaration.