Healthcare AI Safety Gap · Don't break patients
    Peer-reviewed / official9 Feb 2026·UK

    AI chatbots no better than a web search for self-diagnosis

    Oxford / Nature Medicine study of ~1,300 people; users did worse identifying conditions.

    <35%
    of users identified the right condition

    What the research found

    In a randomised study of about 1,300 UK participants, people using AI chatbots to self-diagnose identified the relevant condition less than 35% of the time — no better than using an internet search. Strikingly, the models performed better on their own than in the hands of a worried person.

    Why it matters for providers

    The tool sold as widening access does not reliably help real patients make better decisions — and gets worse under real-world stress.

    Original source
    Oxford / Nature Medicine
    Read the full report ↗
    Peer-reviewed or official source — the most reliable tier.