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

    AI scribes invent and omit content in clinical notes

    Validated eval: 1.47% hallucination and 3.45% omission across 12,999 sentences.

    1.47% / 3.45%
    hallucination / omission rate

    What the research found

    A validated evaluation of an ambient AI scribe across 12,999 clinician-annotated sentences found a 1.47% hallucination rate (fabricated content) and a 3.45% omission rate (missing critical information). Even low rates mean errors in thousands of notes at health-system scale.

    Why it matters for providers

    Ambient scribes are the fastest-spreading clinical AI, so this is the most load-bearing everyday risk — quiet fabrication and omission entering the record.

    Original source
    Frontiers in AI / npj Digital Medicine
    Read the full report ↗
    Peer-reviewed or official source — the most reliable tier.