There are two things you can do with a clinician’s spoken description of a case. You can transcribe it verbatim and format it. Or you can use it as raw material to produce a structured clinical note.

The two objects look alike. They don’t have the same value.

What a clinical note is

A clinical note is a document intended for a reader who wasn’t in the room. That reader may be:

  • The clinician themselves, three months later, picking the patient back up
  • An on-call colleague seeing the patient for the first time
  • An emergency department that must decide in minutes
  • An expert in a medico-legal proceeding

That reader needs to know, in thirty seconds of reading: why this patient is consulting, what state they’re in, what was decided, and what risks to monitor. They don’t need to know the order in which the clinician spoke, what hesitations occurred during the history, or which exact words were said.

A clinical note is therefore, by construction, a judged synthesis, not a faithful transcript.

What distinguishes a good note from a well-presented transcript

A well-presented transcript can look like a note. Let’s put an example side by side. Same clinical information, two formats.

Formatted transcript

So the patient I’m seeing in second consultation, Mr. Dupont, 42, he comes because he says he’s not doing well, he’s been sleeping poorly for three weeks, he doesn’t really feel like doing anything, he’s lost 3 kilos, he talks about not being here anymore but he says he wouldn’t do anything. He’s drinking a bit more, like 3-4 beers in the evening, before it was 1-2. His father died two months ago. He’s already on Ativan 1mg at bedtime that I prescribed last time. I think we’re more on a major depressive episode, post-bereavement, to monitor for risk. I’ll start an antidepressant, Venlafaxine 75, and see him again in two weeks.

This version is readable. It contains all the information. Yet it’s unusable as a clinical note: the reader must re-extract everything themselves.

Structured clinical note

Reason for consultation: Reassessment of post-bereavement depressive episode, 2nd consultation.

History: Father’s death two months ago. Depressive symptomatology for 3 weeks: insomnia, anhedonia, 3 kg weight loss. Increased alcohol consumption (3-4 beers/evening, vs. 1-2 previously).

Mental status exam: Depressive mood, mild psychomotor slowing. Passive death ideation, no plan or intent (“not being here anymore”). No psychotic features. Morbid awareness preserved. Help-seeking present.

Suicide risk: Low to moderate. Passive ideation, no plan, no history, help-seeking present. Aggravating factors: recent bereavement, rising alcohol use. Verbal safety contract accepted.

Clinical impression: Major depressive episode, post-bereavement, with rising alcohol comorbidity.

Plan: Start Venlafaxine 75 mg/day. Continue Ativan 1 mg at bedtime. Alcohol reduction discussed. Follow-up in 2 weeks. Emergency guidance reviewed.

Same information, seven times more useful. A colleague reads this note in twenty seconds and knows what to do.

The difference isn’t in formatting. It’s in what the writer decided to keep, how they categorized it, and how they formulated it in clinical terms. That’s clinical judgment, not layout.

What an AI tool can do — and should do

A good clinical dictation tool performs three transformations:

  1. Structure detection. Recognizing what belongs to the reason, the history, the mental status exam, the plan, and placing each element in the right section. This requires knowledge of the standard psychiatric note format, not knowledge of the patient.

  2. Register translation. Moving from informal to clinical, where the correspondence is unambiguous. “He’s agitated” → “psychomotor agitation.” “He’s flat” → “affective blunting.” “He’s all over the place” → “flight of ideas.” This cleans up the note without changing the meaning.

  3. Cleanup of spoken material. Removing “uhs,” repetitions, reformulations, false starts. Rebuilding complete sentences from a spoken stream. Pure reshaping.

These three operations strictly preserve clinical content. The tool adds nothing. It structures.

What an AI tool should not do

The critical line is this: the tool must never infer, complete, or interpret. Three examples of drift:

Diagnostic inference. The clinician dictates “he’s anxious, he’s not sleeping, he has muscle tension.” The tool must not write “Generalized Anxiety Disorder” unless the clinician explicitly named it. Diagnosis is a medical act. It’s made by the clinician. The tool has no business formulating it in their place.

Inference from a medication. The clinician says “I’m starting Quetiapine 50 mg in the evening.” The tool must not write “for the psychotic disorder” because quetiapine can be prescribed for insomnia, anxiety, mood stabilization, many other indications. Only the clinician knows the indication.

Completing empty sections. If the clinician doesn’t mention suicide risk, the tool must not write “no suicidal ideation expressed” by default. It must omit the section. A clinical note that falsely claims an assessment was done is dangerous.

Upgrading uncertainty. The clinician says “I wonder if there isn’t a bipolar spectrum.” The tool must not turn this into “suspected bipolar disorder” in the note body. It must preserve the doubt.

Adding undictated recommendations. The clinician says “I’m starting metformin.” The tool must not add “endocrinology consultation” on its own.

The single rule: everything in the note must have been explicitly dictated by the clinician. No plausible supplements, no logical extensions, no standard formulas.

Why this is hard to maintain

Large language models are trained to complete and smooth over. That’s their instinct. Asking an LLM to structure clinical dictation amounts to asking it to hold back, which is counter-nature for it.

A serious tool for clinicians compensates upstream with strict, repeated system instructions that prioritize fidelity over completeness. It’s never perfect. A clinician must always review their note before copying it into the record. But the difference between a tool designed to be faithful and one designed to be impressive is visible from the first notes.

The simple test

To evaluate a clinical note tool, a pragmatic test: deliberately dictate an incomplete note, explicitly omitting certain sections (e.g., don’t speak about suicide risk, don’t give a follow-up plan).

  • If the tool returns a note with an empty or absent risk section, it is faithful.
  • If the tool fills the section with a generic formula (“no reported suicidal ideation,” “follow-up to be organized”), it has invented. It is dangerous.

This test takes five minutes. It separates tools that help clinicians from tools that expose them.

Conclusion

The clinical note is an object of judgment, not of layout. What makes it useful isn’t the faithful transcription of the clinician’s speech — it’s the selection, structuring, and formulation the clinician performs during dictation.

An AI tool must serve this work. Structure what is said. Don’t add what isn’t said. Let the clinician remain the sole source of clinical judgment in the note.

It’s less spectacular than a tool that “writes the note for you.” It’s also the only way to produce a note that holds up in a record and in front of a clinical reader.