Across generations, sizes, and types, large language models poorly report self-confidence in gastroenterology clinical reasoning tasks

Key Points

Models demonstrated poor self-estimation of certainty despite varying accuracy levels.
Even the best-performing models showed significant overconfidence with Brier scores between 0.15 and 0.2.
Analysis involved 48 large language models tested with 300 gastroenterology board exam-style questions.
Human oversight is essential, as current models cannot communicate uncertainty effectively.

Abstract

This study evaluated confidence calibration across 48 large language models (LLMs) using 300 gastroenterology board exam-style multiple-choice questions. Regardless of accuracy, all models demonstrated poor self-estimation of certainty. Even the best-calibrated systems (o1 preview, GPT-4o, Claude-3.5-Sonnet) exhibited substantial overconfidence (Brier scores 0.15–0.2, AUROC ~ 0.6). Models maintained high confidence regardless of question difficulty or response correctness. In their current form, LLMs cannot be relied upon to communicate uncertainty, and human oversight remains essential for safe use.

Bookmark

View Full Paper

Cite This Study

Naderi et al. (Thu,) studied this question.

synapsesocial.com/papers/69a767febadf0bb9e87e3309 https://doi.org/https://doi.org/10.1038/s44355-026-00053-3

Bookmark

View Full Paper