Healthcare AI Safety Gap · Don't break patients
    Reputable report / newsJun 2026·UK / Intl

    One false sentence flips a medical AI from right to wrong

    Oxford's MedMisBench: mean accuracy fell 71%→38% across 11 models under a misleading cue.

    71% → 38%
    mean accuracy after one false sentence

    What the research found

    MedMisBench tested 11 leading models — including GPT-5, Gemini and Claude — on roughly 11,000 real medical questions. When a single plausible-but-false sentence was injected alongside a question the model had answered correctly, mean accuracy collapsed from 71% to 38%, and more than half of correct answers flipped to wrong. The models had the knowledge; the context steered them away from it.

    Why it matters for providers

    Real patients and clinicians rarely give clean prompts. This shows any AI can be pushed to a confidently wrong answer by one misleading detail — the exact condition of everyday use.

    Original source
    Oxford et al. — MedMisBench (arXiv / bioRxiv preprint)
    Read the full report ↗
    A reputable institutional report or news outlet.