Healthcare AI Safety Gap · Don't break patients
    Reputable report / news2025·US / Intl

    LLMs often fail to challenge users' delusions

    Studies warn of 'AI psychosis' risk for vulnerable and mental-health users.

    delusion risk
    models reinforce rather than de-escalate

    What the research found

    Emerging simulation studies find large language models frequently fail to challenge users' delusional statements, sometimes reinforcing them rather than de-escalating — a specific risk in mental-health and companionship use.

    Why it matters for providers

    Unsupervised AI can amplify, rather than defuse, a crisis for the most vulnerable users. Treat as emerging evidence.

    Original source
    JMIR and related research
    Read the full report ↗
    A reputable institutional report or news outlet.