Explainable AI changes dermatological diagnosis decisions differently depending on who is holding the screen, according to a new study published in Nature. The same visual explanation that helps a clinician can mislead a layperson. That divergence is the finding, and it matters for anyone shipping AI triage tools into clinics or consumer apps.

The paper, titled "Divergent impacts of explainable AI for dermatological diagnosis on clinicians versus lay people," examines what happens when artificial intelligence systems that flag skin conditions also show why they reached a conclusion — typically through heatmaps or saliency overlays highlighting lesion regions of interest.

Why does this matter for dermatology?

Skin-disease AI has moved fast. Algorithms now classify images of pigmented lesions, inflammatory disease and tumors with accuracy that rivals specialists in narrow benchmarks. But a raw prediction isn't enough in a clinical workflow. Regulators, ethicists and clinicians have pushed for "explainability" — visual or textual justifications attached to each output — on the theory that transparency builds trust and catches errors.

The Nature study tests that theory. And the theory doesn't fully hold.

The core result: explanations do not affect clinicians and lay people the same way. The two groups respond divergently to identical explanatory content. What functions as a decision aid for one audience can shift behavior in a different — and potentially less desirable — direction for the other.

What does the divergence mean in practice?

Think about who actually encounters these tools. A dermatologist may use AI as a second reader, weighing the model's highlighted regions against their own clinical assessment of a lesion's morphology. A patient using a smartphone app has no such scaffold. They see a confident overlay on a photo of their own skin.

If the explanation increases a clinician's appropriate scrutiny, good. If the same overlay inflates a layperson's confidence in a prediction — right or wrong — that's a different story. The study's framing suggests exactly this kind of split: the impact of explainability is audience-dependent, not universal.

The takeaway is plain: an explanation designed for experts is not automatically safe for consumers, and vice versa.

Where does this leave AI skin apps?

Most consumer dermatology apps and clinical decision-support tools share a common engineering assumption — that explainability is an unqualified good. The Nature findings complicate that. Developers may need audience-specific explanation formats: richer, caveat-laden justifications for clinicians; simpler, calibrated and possibly more conservative framing for lay users.

For dermatologists evaluating AI vendors, the study offers a practical screening question. Don't just ask whether the system explains itself. Ask who the explanation was designed for, and whether evidence shows it helps — or distorts — decisions for that specific audience.

As AI-assisted skin diagnosis moves from journal benchmarks into waiting rooms and app stores, expect explanation design to become as contested as the algorithms themselves.