Picture two people looking at the exact same AI-generated medical explanation — a dermatologist and a worried patient searching symptoms online. One catches the AI’s mistake instantly. The other might not catch it at all. That’s the core problem a new study just exposed: health AI interfaces built the same way for everyone can quietly put non-experts at risk.
The Study Behind the Warning
Same Tool, Different Outcomes
Published in Nature Medicine, the research focused on AI-powered dermatological diagnosis — a field where AI already assists clinicians and increasingly reaches everyday patients through AI search tools. Lead author Orson Xu, of Columbia University’s Department of Biomedical Informatics, found that explanations meant to build trust in AI don’t work the same way for everyone.
A clinician typically forms their own diagnosis first, then checks the AI’s reasoning against their training. A weak explanation gets caught quickly. A non-expert, however, often forms their opinion from the AI’s explanation — meaning a confident-sounding but flawed rationale can steer them straight toward the wrong answer.
Explainability Isn’t One-Size-Fits-All
This is the heart of the issue: the same explainable AI feature can act as a safety check for one user and a source of error for another. The researchers argue against treating explainability as a fixed, universal interface component.
What This Means for AI Interface Design
The Automation Bias Problem
Marzyeh Ghassemi of MIT’s Department of Electrical Engineering and Computer Science noted that good AI systems can genuinely improve outcomes in healthcare — but only when balanced against something called algorithmic deference, where users lean too heavily on the machine’s judgment instead of their own.
Both AI outputs and explainability tools can trigger automation bias, an anchoring effect where people over-trust a system simply because it sounds authoritative. Designers, Ghassemi says, need to actively account for this rather than assume transparency alone solves the trust problem.
Designing for the Whole Spectrum of Users
The takeaway isn’t that AI explanations are bad — it’s that user expertise in AI should shape how those explanations are built and delivered. A dermatologist reviewing a lesion needs different interface cues than a patient using a symptom-checker app at home.
This means healthcare AI products may need tiered explanation systems: deeper technical detail for professionals, and clearer guardrails or uncertainty flags for the general public.
Conclusion — Trust Needs Context, Not Just Transparency
As AI in healthcare becomes more common, this study is a timely reminder that good intentions — like making AI “explainable” — aren’t enough on their own. Interfaces must be built with the user’s expertise in mind, not a one-size-fits-all design. As AI tools reach more patients directly, getting this right isn’t optional — it’s essential for safe, effective care.
