Who signed the note the model wrote? Attribution for ambient AI scribes
The ambient scribe writes the note. The clinician clicks attest. At that instant a model's output becomes a legal medical record, a billing artefact and a discovery exhibit. The record shows a signature. It does not show a reading — and in a malpractice matter, the difference between those two is the case.
Who signed the note the model wrote?
A clinician, in a click that is indistinguishable from one applied to a note they composed themselves. Ambient scribes now draft a large share of clinical documentation, and nothing in the record distinguishes a clinician who read every line from one who signed forty notes in four minutes.
- Every AI governance framework in healthcare requires a clinician to review model-generated documentation. None specifies what review looks like as evidence.
- The measurable proxy is notes attested per minute. It is computable from data your EHR already holds, and it is the first number opposing counsel will ask for.
- Binding the attestation to a hash of the exact rendered note, plus a diff fingerprint against the model draft, converts an unfalsifiable policy claim into a testable artefact.
Part of Hospital and clinical identity
The measurement that already exists in your data
Before discussing mechanism, establish the denominator. Your EHR records note creation time, attestation time and clinician identifier for every encounter note. Three columns. Export them for one service line for one month.
Compute attestations per clinician per session, and the interval between successive attestations. Report the distribution rather than the mean; the mean will be inflated by clinicians who attest one note at a time between patients, which is not the behaviour under question.
| Median interval | Plausible reading | What it supports in a deposition |
|---|---|---|
| Under 5 seconds | Batch attestation of a queue | Nothing. The record documents a session, not a review. |
| 5 to 20 seconds | Scanning headings and the assessment | Partial review of structured sections only. |
| 20 to 90 seconds | Reading the narrative | Defensible for a typical encounter note. |
| Over 90 seconds | Reading and editing | Strong, if the edits are captured. |
There is no published national figure for this and this article does not assert one. The point is that the number exists inside your organisation, has never been computed, and will be computed eventually by somebody whose interests are not aligned with yours.
Why the current record cannot answer the question
The attestation is a user interface event written into the EHR by the EHR. Three specific gaps follow, and each has been the hinge of a documentation dispute in the pre-AI era already.
- The bytes are not bound. Note text can be re-templated, macro-expanded or post-processed between display and storage. Nothing ties the attestation to the exact characters the clinician saw.
- The draft is not retained comparably. Most deployments discard or archive the model's original output separately, so the question what did the clinician change requires forensic reconstruction rather than a lookup.
- Presence is inferred. The attestation is attributed to a session. On shared clinical workstations, session attribution is itself unreliable.
The override rate is the honest signal
Here is the uncomfortable statistical point. If a clinician reviews carefully and the model is good, the override rate will be low. If a clinician does not review at all, the override rate will be zero. Those two states produce nearly identical records, and the second is indistinguishable from the first by any measure currently collected.
That is why override rate on its own is a weak governance metric, and why it is dangerous to celebrate a low one. What separates the two states is not the outcome but the process, and the process is exactly what is not recorded.
A metric that looks identical whether the control worked or was skipped is not a control metric.
The Clinical Attestation Statement
The fix is to make the attestation cover what it claims to cover. A canonical statement, signed at the moment of attestation by the clinician's hardware credential, carrying:
| Field | Content | What it establishes |
|---|---|---|
note_hash | Hash of the exact rendered note text at attestation | The clinician attested to these characters, not a later version |
draft_hash | Hash of the model's original output | The draft can be compared later without retaining two full copies indefinitely |
diff_fingerprint | Compact summary of edits: sections touched, characters changed | Distinguishes reading-with-edits from clicking |
model_id | Scribe model identifier and version | Ties the draft to a specific model generation for later evaluation |
dwell_ms | Time the note was displayed before attestation | Removes the need for forensic timeline reconstruction |
clinician | Hardware-backed assertion, user verification required | Names a human rather than a session |
No clinical content leaves the perimeter. The receipt carries hashes and a fingerprint, which is enough to prove correspondence and insufficient to reconstruct the note.
What this does to the deposition
Currently, counsel defending an AI-assisted documentation claim has a policy, a training record and a timestamp. Opposing counsel has a theory that nobody read the note, and no way to test it either — which sounds balanced but is not, because the burden of explaining a fifteen-second attestation interval falls on the clinician.
With a receipt, the factual question resolves before argument. Either the record shows a two-minute dwell with edits across three sections, or it shows a four-second dwell with no edits. Both outcomes are better than an unresolvable dispute, including the unflattering one, because the unflattering one is discoverable anyway through timestamps.
Scoping this sensibly
Not every note warrants it. Reasonable scoping by risk:
- Notes that support a billing level where documentation is the sole evidence of service complexity.
- Notes containing a diagnosis that drives a coverage or authorisation decision.
- Any note in a specialty with elevated malpractice frequency, decided by your carrier's data rather than intuition.
- Any note where the model generated an assessment or plan, as opposed to transcribing history.
That is a minority of total volume in most deployments, which is what makes the control affordable.
A note on the clinicians
Ambient documentation is, by a wide margin, the most popular clinical technology deployment of the decade, because it gives clinicians their evenings back. Nothing here argues against it. The argument is that a technology this valuable deserves governance that can survive contact with a courtroom, and that building the evidence now is cheaper than reconstructing it later.
What the record should distinguish
| State | Should be recorded |
|---|---|
| Clinician composed directly | Yes |
| Model drafted, clinician edited | Yes — with what changed |
| Model drafted, signed unedited | Yes — and it is the one that matters |
Provenance is becoming the question that quality review, malpractice defence and payer audit all converge on. A record that says which of the three applies is materially more defensible than one that shows only a signature.
Objections and honest limits
“The clinician is responsible regardless of who drafted it.” They are, and that is precisely why the record should show what they were responsible for. Undifferentiated attribution exposes the careful clinician to the same inference as the careless one.
“Editing metrics will be used against clinicians.” A real risk, and the reason to record provenance rather than a productivity metric. Whether the text was generated is a fact about the document; how fast someone worked is not the same thing.
Recording note provenance
- Record the provenance state. Composed, edited, or signed unedited.
- Record what changed, if anything. A digest before and after editing.
- Bind the signature to the final text. So later alteration is detectable.
- Tier the attestation. Full review for medication, allergy and diagnosis changes.
Terms used here
- Ambient scribe
- A system generating clinical documentation from a recording of the encounter.
- Provenance
- Whether text was composed, edited or generated — a fact about the document rather than about the clinician.
- Attestation
- The clinician's signature converting a draft into the legal medical record.
Frequently asked questions
Does the receipt contain patient information? No. It carries one-way hashes of the note text and the model draft, plus a compact diff fingerprint. Nobody can reconstruct clinical content from it.
Will clinicians accept another click? They already perform the attestation click. The change is that it produces a signed statement rather than a database row; on a device with biometric unlock the additional gesture is roughly one second.
What if the model draft was perfect and no edits were needed? Then the receipt shows a genuine dwell time with no edits, which is a defensible record. The problem today is that this case and the no-review case are indistinguishable.
Is this required by any regulation? No current regulation prescribes this artefact. Frameworks require human oversight of AI-generated clinical content without specifying its evidentiary form, which is precisely the gap being filled.
Why record provenance? Because quality review, malpractice defence and payer audit are all converging on the question, and an undifferentiated signature answers none of them.
Does this expose clinicians to editing metrics? It should record provenance, not productivity. Whether text was generated is a fact about the document; how fast someone worked is not.
Which notes need full review? Those changing medication, allergies or diagnosis — a minority, which is what makes tiering workable.
Where this fits in Manav
Manav binds the clinician to the exact order, note or session being authorised, with a gesture short enough for the bedside and a record an investigator can verify without the EHR.
Sources and further reading
- Federal and state requirements on physician attestation and authentication of medical record entries.
- Payer documentation and audit standards for clinician attestation of encounter notes.
- NIST AI Risk Management Framework
- Regulation (EU) 2024/1689 (AI Act) — consolidated text
- RFC 8785 — JSON Canonicalization Scheme
- AMA — augmented intelligence in medicine
- 45 CFR §164.312 — HIPAA Security Rule technical safeguards