New Benchmark Reveals LLMs Fail to Preserve Diagnostic Uncertainty in Clinical Text

A new study from researchers Hongbo Du, Zixin Lu, and Jiaming Qu tackles an underexplored risk in clinical AI: whether large language models (LLMs) correctly preserve diagnostic uncertainty when summarizing or revising clinical text. In medical documentation, phrases like "possible pneumonia" vs. "definite pneumonia" convey evidence strength and directly influence follow-up testing and treatment decisions. Altering these uncertainty expressions can completely change clinical meaning, yet most evaluation benchmarks only measure fluency and coherence. To address this gap, the team constructed a benchmark of 1,200 clinical documents containing 9,184 uncertainty annotations across five levels, then tested three unnamed LLMs on their ability to maintain the original uncertainty cues.

The results are concerning: LLMs preserved original uncertainty cues less than half the time and particularly struggled with subtle differences between adjacent levels (e.g., "possible" vs. "probable"). This failure mode is invisible to standard evaluation metrics that focus on grammatical correctness or readability. The study highlights a critical safety gap for AI in clinical settings—deploying LLMs without evaluating uncertainty preservation could lead to downstream diagnostic errors or inappropriate treatment decisions. The authors call for new evaluation standards that explicitly test uncertainty handling before LLMs are integrated into clinical workflows, making this a must-read for healthcare AI practitioners and regulators.