Rapid Clinical Updates: Diagnosis in the Era of AI: The Good, The Bad, and The Ugly-Slides

Slides

Pdf Summary

This “Rapid Clinical Updates” session, moderated by Ethan Molitch-Hou, features Daniel Restrepo (MGH/Harvard) and Ethan Goh (Executive Director, Stanford ARISE) discussing clinical diagnosis in the AI era—“the good, the bad, and the ugly.” The talk emphasizes that modern large language models (LLMs) can perform impressively on standardized exams and curated diagnostic reasoning benchmarks, but they are not poised to replace clinicians soon. Instead, they are likely to <em>enable</em> clinicians if used carefully, while requiring continued vigilance and strong clinical fundamentals.

<strong>The “Good” (promises/use cases):</strong> LLMs can accelerate knowledge, guideline, and reference retrieval; improve chart/EMR summarization (e.g., summarizing prior hospitalizations, medication changes, or frequency of presentations); and provide “cognitive second opinions,” especially for complex cases, unfamiliar presentations, or multimorbidity. They may assist with hypothesis generation (“what am I missing?”), tailored differential diagnosis generation, and potentially EMR-informed base rates or early-warning nudges about cognitive bias.

<strong>The “Bad” (limitations):</strong> Model performance depends heavily on the quality of clinical inputs; real-world medicine is noisy compared with “crisp, curated” case vignettes used in evaluations. Therefore, history-taking and physical examination remain essential, and inaccurate/incomplete data can lead to incorrect AI outputs. Hallucinations persist—especially for esoteric questions—making reference-checking critical and discouraging reliance on open-access models for clinical care.

<strong>The “Ugly” (perils):</strong> AI can amplify cognitive and racial biases (e.g., suggestibility, confirmation, availability, framing). Another major concern is cognitive deskilling: overdependence may erode clinicians’ diagnostic skills and trainees may “outsource” reasoning. The session stresses role-modeling appropriate AI use and maintaining fund-of-knowledge and pattern recognition as core competencies.

Keywords

clinical diagnosis

large language models

AI in healthcare

diagnostic reasoning

EMR summarization

clinical decision support

differential diagnosis generation

LLM hallucinations

cognitive bias amplification

clinician deskilling