Why your AI scribe should never write what wasn't said

July 5, 2026 · The Visit to Chart team

Large language models have a well-documented failure mode: when they don't know something, they can produce something plausible instead. In a chatbot, that's an annoyance. In a clinical note, it has a different name — fabricated documentation — and it can follow a patient through every future encounter, medication decision, and records request.

How hallucination shows up in clinical notes

The dangerous cases aren't outlandish; they're boring and plausible. Note templates train models on what notes usually contain, so an unguarded model tends to fill in what's usually there:

• A "normal" physical exam that was never performed — common in telehealth, where much of a traditional exam can't be performed.
• Vitals or labs that were never mentioned, borrowed from the statistical average of a thousand similar notes.
• Risk-assessment language — the most consequential text in some specialties — templated in rather than grounded in what the patient actually said.
• A plan item ("patient agrees to follow up in 2 weeks") that sounds right but never happened.

Reviews of ambient-scribe deployments consistently list accuracy, governance, and overtrust as the open problems alongside the time savings. The technology is genuinely useful; it just needs to be engineered like it will occasionally be wrong — because it will be.

What real guardrails look like

When you evaluate any AI documentation tool, the question isn't "is it accurate?" — everything is accurate in the demo. The question is what happens on the bad day. Look for four specific design decisions:

1. A no-invention rule. The drafting system should be explicitly built never to generate exams, vitals, labs, or risk language that isn't supported by the transcript. Not "rarely" — never, as a design constraint.
2. Flag, don't fill. When something usually belongs in the note but wasn't said in the visit, the right behavior is a visible gap handed to the physician ("not addressed in visit — complete if performed"), not a guess.
3. A second pass that checks the first. The strongest pattern is adversarial: after drafting, a separate AI pass compares every statement in the note against the transcript and flags anything it can't support. The physician sees exactly which sentences to scrutinize instead of proofreading blind.
4. An honest signature workflow. The note carries an AI-assistance disclosure, and nothing is final until the physician reviews and signs. If the product makes it easy to sign without reading, its incentives are wrong.

The physician stays the author — and that's the point

None of this diminishes the value of the tool. A scribe that drafts faithfully and flags honestly turns review into minutes of verification instead of an evening of composition. The goal was never to remove the physician from documentation — it was to remove the typing. The judgment, the signature, and the authorship were never the burden.

Sources

• Ambient AI Medical Scribes: Efficiency Gains, Burnout Uncertainty, and Governance Risks (IHS)
• Barriers and opportunities of scaling ambient AI scribes (npj Digital Medicine)
• Ambient Scribes Lower Physician Workload but Introduce New Challenges (Docwire News)
• Ambient AI Scribes to Alleviate the Burden of Clinical Documentation (NEJM Catalyst)