Why every sentence in an AI note should cite its source

Large language models are very good at writing a progress note that sounds right. That is exactly the problem. A note that sounds right gets signed. And once it's signed, it's the clinician's attestation — every sentence, including the one the model invented.

Ambient documentation has moved quickly from pilot to default in a lot of health systems, and the evidence on time savings is encouraging. The evidence on accuracy is more mixed. Published evaluations of AI scribes consistently find two kinds of error: omissions, where something clinically important was said but didn't make it into the note, and fabrications — often called hallucinations — where the note states something that wasn't said at all. Fabrications are rarer, but they're the ones that end up in court, in an audit, or in a client's records request.

Why fluency is the risk#

A human-written note with a mistake usually looks like it has a mistake: a typo, a half-finished thought, a copied block that doesn't fit. An AI-written mistake looks exactly like the sentences around it. The model fills a plausible gap — a medication dose, a homework assignment, a denial of suicidal ideation — with a plausible answer.

Behavioral health makes this worse in three ways. Sessions are long and conversational, so there's more room for the model to summarize loosely. The most consequential content — risk, substance use, safety planning — is often said briefly and indirectly. And clinicians review notes at the end of a long day, when a fluent paragraph is most likely to get a skim and a signature.

What sentence-level evidence means#

The fix isn't a better disclaimer. It's making every claim in the note checkable in one click. In practice that means each sentence carries a link to its source:

• Transcript: the exact moment, by timestamp and speaker, where it was said.
• Chart: the record it came from — the treatment plan, the medication list, last week's PHQ-9 — clearly labeled from chart rather than from session.
• Clinician: content the clinician added or edited, attributed to them.

Then a second, independent pass verifies each sentence against its cited source and grades it: supported, partially supported, unsupported, or contradicted. Anything short of supported is highlighted before the clinician signs, with the transcript excerpt beside it.

Why this changes review#

Without evidence, review means re-reading everything and trusting memory. With evidence, review becomes triage: supported sentences can be skimmed, flagged sentences get attention. Clinicians spend their scarce attention where it matters. That's the honest version of 'saves time' — not faster signing, but faster correct signing.

It also changes what happens later. When a supervisor, an auditor or a client asks 'where did this come from?', there's an answer. The transcript is retained only as long as your organization's policy allows, but the provenance — which sentence came from where, who edited what, which check passed — remains part of the audit trail.

Design choices that make it work#

Citations are only useful if they're honest. A few rules we've found necessary:

1. 1No citation, no sentence. If the drafting model can't point to a source, the sentence is either dropped or explicitly labeled as unsupported. Generic filler ('client tolerated session well') doesn't get a pass.
2. 2Verification uses a different prompt and, where possible, a different model than drafting, so it isn't grading its own homework.
3. 3Chart content is labeled as chart content. Prior-visit facts carried into a note are marked so nobody mistakes them for something said today.
4. 4Risk language is held to a higher bar. Statements about suicidal ideation, homicidal ideation, abuse or substance use require a direct transcript match, or they're flagged.
5. 5The clinician is the author. Edits are attributed. Nothing is submitted until a human signs, and the system never signs on anyone's behalf.

What it doesn't solve#

Evidence doesn't fix omissions on its own — a sentence that was never written can't be flagged as unsupported. That's why Margent pairs citations with a completeness check against the note template and the organization's rules: if risk was discussed but no risk section exists, or a treatment-plan goal was worked on but not referenced, that's a finding too. And evidence can't make a poor recording good; crosstalk, distance from the microphone and interpreters all reduce transcript quality, and the confidence indicators say so.

Questions to ask any AI documentation vendor#

• Can I see, for any sentence in the note, exactly where it came from?
• What happens to a sentence the system can't support — is it dropped, flagged, or written anyway?
• Is verification done by the same model call that wrote the note?
• How are risk statements (SI, HI, abuse, substance use) handled differently from the rest of the note?
• Who can see the edit history, and does it survive after the transcript is deleted?
• Do you publish how often notes are edited before signing, and how you measure accuracy?

If you're evaluating AI documentation tools, ask any vendor to show you a draft with a deliberate error in it. Watch how long it takes to find. Then try ours — no signup, a seeded CBT session with an unsupported sertraline sentence planted in it.