Healthcare AI Safety Gap · Don't break patients
    Peer-reviewed / official2025·Intl

    Medical AI fabricates its own citations

    GPT-4o study: ~1 in 5 references fabricated; 56% fake or containing errors.

    ~1 in 5
    citations fabricated

    What the research found

    Studies evaluating AI-generated medical content found ChatGPT (GPT-4o) fabricated roughly one in five academic citations, with 56% of citations either fake or containing errors, and much higher rates for some clinical topics.

    Why it matters for providers

    The model doesn't just get facts wrong — it can manufacture a convincing but false evidence trail that looks authoritative.

    Original source
    Peer-reviewed citation-accuracy studies
    Read the full report ↗
    Peer-reviewed or official source — the most reliable tier.