Too agreeable to be accurate? Sycophancy and diagnostic instability of large language models in medical diagnosis

Key Findings

• — Across all neutral model-case pairs, accuracy fell from 51.8% to 42.2% after the prompt 'Are you sure?', an accuracy-changing flip rate of 32.8%.
• — Correct-to-incorrect transitions numbered 544 against 198 incorrect-to-correct, so the challenge broke about 2.7 right answers for every one it fixed.
• — Claude Sonnet 4 had the highest flip rate (58.6%) and the largest post-challenge accuracy loss (-36.4 percentage points).
• — GPT-5 was the most stable, moving from 74.4% baseline to 75.3% after the challenge.
• — Specialty framing mattered far less: adjacent-specialty prompts added 2.8 percentage points, differential-specialty prompts removed 1.2 to 1.8.
• — 28,800 responses from 120 vignettes (40 MultiCaRe clinical narratives, 80 MedMCQA exam-style cases) across ten models; the authors recommend that diagnostic evaluations report harmful flips and stability, not only first-pass accuracy.

Methodology Notes

Diagnostic correctness was graded by an LLM-as-a-judge, which is the main methodological limit; exam-style MedMCQA cases inflate baseline accuracy relative to the narrative MultiCaRe cases, and the authors report that case source strongly affected results. No clinicians or patients in the loop. Published online ahead of print 2026-09-04 in Artificial Intelligence in Medicine 182:103516, DOI 10.1016/j.artmed.2026.103516; sciencedirect 403s from this box, so the record was verified from Crossref metadata and the full PubMed abstract (PMID 42721584).

Sources