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

Video Transcription

Video Summary

The session is a “rapid clinical update” on diagnosis in the era of AI—its benefits, limitations, and risks for hospitalists. Moderator Ethan Mollick (University of Chicago) introduces speakers Daniel Restrepo (MGH/Harvard; clinical reasoning and educator) and Ethan Goh (Stanford ARISE; AI research leader).  <br /><br />Restrepo focuses on practical diagnostic use cases for large language models (LLMs): faster knowledge retrieval (e.g., summarizing dense guidelines, quickly finding data with citations), EMR/chart summarization, and using AI as a “cognitive second opinion” for complex or unfamiliar cases. He demonstrates how prompting affects outputs using a case that turned out to be multifocal pyogenic liver abscesses from smoldering diverticulitis; when key historical details are included, AI prioritizes the correct diagnosis.  <br /><br />He then outlines major limitations and “ugly” pitfalls: AI depends on clinician-provided data (garbage in/garbage out), real-world histories are messy, hallucinations still occur (especially with esoteric questions), models can inherit human biases, and overreliance may cause cognitive de-skilling—particularly concerning for trainees.  <br /><br />Goh emphasizes we’re early in understanding failures, models change rapidly, and privacy/workflow drive EHR integration. Both recommend experimenting with tools cautiously, verifying sources, avoiding PHI in public models, and using AI mainly to augment—not replace—clinical reasoning.

Keywords

AI in clinical diagnosis

large language models (LLMs)

hospitalist workflow

clinical reasoning support

EHR/EMR chart summarization

prompt engineering in medicine

diagnostic decision support

AI hallucinations and bias

privacy and PHI safeguards