MedAI Digest

LLMs Are Vulnerable to Cognitive Bias Attacks on Radiology Exams


Listen Later

Researchers discovered that large language models tested on radiology board-style questions are highly vulnerable to cognitively biased prompts, with accuracy plummeting by up to 45% under certain attack conditions. Original paper: Cognitively Biased Prompt Effects on Large Language Model Accuracy for Radiology Board-Style Examination Questions. — Radiology. Artificial intelligence. 10.1148/ryai.250585 📄 Read the article
...more
View all episodesView all episodes
Download on the App Store

MedAI DigestBy MedAI Digest