Perspective Video Interview: Managing Uncertainty

Why It Matters

Embedding explicit uncertainty signaling in clinicians and AI safeguards high‑stakes decisions, reducing diagnostic errors and fostering trust in digital health tools.

Key Takeaways

• Clinicians must disclose knowledge gaps, but training lacks uncertainty competency.
• AI models often confabulate instead of admitting ignorance, posing safety risks.
• Introduce ‘I don’t know’ as core entrustable activity for clinicians and AI.
• Benchmarking clinical vignettes can assess uncertainty expression in trainees and AI.
• Embedding epistemic humility improves decision‑making in high‑stakes bedside care.

Summary

In a NEJM perspective interview, Steven Morsy and Dr. Raja Ali Abdul Nure discuss how uncertainty permeates modern medicine and why clinicians and AI systems must learn to vocalize it. The conversation frames uncertainty as both factual—diagnostic probabilities—and environmental—team dynamics or equipment location—highlighting its omnipresence in daily practice. The duo argues that medical education, built on competency‑based curricula, has never codified the act of saying “I don’t know” as a measurable behavior. They propose an entrustable professional activity (EPA) that makes epistemic humility observable and assessable, bridging the gap between recognizing knowledge limits and deliberately shifting to analytical reasoning. Illustrative examples include surgeons physically slowing their hands when faced with ambiguous anatomy and large language models confidently hallucinating drug names, even when presented with a fictitious “Pokemon” medication. Recent benchmark studies that embed falsehoods in clinical vignettes reveal AI’s reluctance to admit uncertainty, underscoring the need for systematic testing. The interview concludes that both training programs and AI developers should adopt uncertainty benchmarks, making “I don’t know” a core competency. Such standards could enhance patient safety, improve shared decision‑making, and set a new bar for trustworthy clinical AI.

Original Description

Raja-Elie Abdulnour, MD, discusses the need for clinicians and AI systems to refrain from expressing undue confidence in their knowledge.

Read the related Perspective article by Andrea Sikora, Pharm.D., M.S.C.R., Leo A. Celi, M.D., M.P.H., and Raja-Elie E. Abdulnour, M.D.: Can AI Say “I Don’t Know”? https://nej.md/4fgdg0D