Stanford study reveals LLM health risks under misinformation
Misinformation in prompts drops LLM accuracy by 7.2% in public health tests
Deep Dive
A study by Chuqing Zhao and Haochen Yang found that in public health applications, misinformation framing degrades LLM accuracy by 7.2 percentage points on average, with prediction flip rates between 9% and 38%. The team tested two perturbation types—misinformation framing and layperson rewriting—and highlighted deployment risks when non-clinical users query health AI systems.
Key Points
- Stanford researchers evaluated LLM robustness in public health scenarios with misinformation framing (MF) and layperson rewriting (LR) perturbations
- MF caused 7.2% accuracy degradation and 9-38% prediction flip rates, while LR caused only 1.4% degradation
- Findings highlight risks of incorrect outputs from misinformation and misinterpretation of informal patient language
Why It Matters
Critical insights for deploying LLMs in healthcare where misinformation and informal language could lead to harmful decisions.