LLMs not yet safe for autonomous clinical triage, paper finds

A new paper argues that while large language models (LLMs) can pass medical licensing exams and assist in clinical tasks, they are not yet safe for autonomous patient triage. The core issue is not a lack of medical knowledge, but the models' inability to safely gather information under uncertainty and identify critical, potentially missed diagnoses. LLMs optimized for predicting probable text may fail to exhibit the necessary behaviors for safe triage, such as broadening differentials, seeking red flags, and escalating concerns when high-harm diagnoses remain unexcluded. AI

IMPACT Highlights critical safety concerns for LLM deployment in healthcare, suggesting a need for further research into information gathering and risk assessment capabilities.

RANK_REASON Academic paper discussing limitations of LLMs in a specific domain. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Autonomous Clinical Decision Support
• Clinical Decision Support
• Hugging Face
• large-language models
• Medical licensing examinations in the United States.
• physician
• Triage

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →