Across generations, sizes, and types, large language models poorly report self-confidence in gastroenterology clinical reasoning tasks
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.
Authors
- Zahra Atf (ORCID: https://orcid.org/0000-0003-0642-4341)
- Peter Lewis (ORCID: https://orcid.org/0000-0002-7844-6200)
- Nariman Naderi (ORCID: https://orcid.org/0000-0003-4820-5497)
- Seyed Amir Ahmad Safavi‐Naini (ORCID: https://orcid.org/0000-0001-9295-9283)
- Ali Soroush (ORCID: https://orcid.org/0000-0002-9485-9662)
- Thomas Savage (ORCID: https://orcid.org/0000-0003-4828-5802)
- Girish Nadkarni
- Mohammad Amin Khalafi
Institutions
- Hospital of the University of Pennsylvania (US)
- Ontario Tech University (CA)
- Research Institute for Endocrine Sciences (IR)
- Shahid Beheshti University of Medical Sciences (IR)
- Icahn School of Medicine at Mount Sinai (US)
Publication Details
- Journal
- npj gut and liver.
- Published
- 2026-02-05
- DOI
- https://doi.org/10.1038/s44355-026-00053-3
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Amgen
- American College of Gastroenterology
- American Gastroenterological Association