Why every Aimé note cites its source

There's a class of AI bug that's tolerable in most software and completely unacceptable in clinical documentation: the hallucination.

A language model will sometimes produce a sentence that sounds right, reads well, and isn't supported by anything the patient said. In a blog post, that's annoying. In a chart, it can be a licensing event.

This is why Aimé cites every clinically meaningful sentence in a note against the specific moment in the session that produced it.

What "citation" means here

When you open a draft note in Aimé, most sentences have a small marker next to them. Click the marker and the audio jumps to the point in the session the sentence came from, with the surrounding transcript highlighted.

The mechanism is simple:

1. The session audio is transcribed with speaker labels and timestamps.
2. When the model drafts a note, it's prompted to anchor each clinical claim to a specific span of the transcript.
3. The draft stores that anchor alongside the sentence.
4. The editor renders the anchors as citations you can jump to.

If the model produces a sentence it can't anchor, the sentence is flagged. You still see it, but you know right away that it's inferred rather than recorded.

Why this matters more than it sounds

A few things become possible when every sentence has a source:

Audits become trivial. When a payer or supervisor asks "where did this observation come from?", you click the citation and play the six seconds of tape.

Editing becomes faster. You're not reading the note in a vacuum. You can verify anything questionable in a few seconds.

Trust becomes the default. Clinicians new to ambient AI almost always start by reviewing every sentence of every note for a week or two. Citation makes that checking mechanical rather than anxious.

What we won't do

A few things are explicitly off the roadmap for this system:

• We won't produce notes without citations. A citation-less draft is not a product feature we're willing to ship.
• We won't train on your session audio. Citations are a product feature, not a data-collection pretext.
• We won't hide flagged-inference sentences. If the model had to guess, you should see the guess and decide.

Why this is a floor, not a ceiling

We don't think citation is some remarkable achievement on our part. We think it's what any AI documentation tool for clinical use should have had from day one. The fact that many don't is a market failure, not a technical one.

If you're evaluating AI scribes, ours or anyone else's, ask the vendor to show you where in the session a specific sentence came from. If they can't do it in two clicks, the tool isn't ready for your chart.