Who Signs an AI-Generated Clinical Note?

A note in a patient record is a legal document. It is the thing a GDC case examiner reads three years from now, when nobody can remember the appointment and the patient’s recollection has hardened into certainty. It is what an indemnifier leans on when deciding whether to defend you. So the question of who signs an AI-generated clinical note is not administrative housekeeping. It is the question of who carries the case.

The answer is uncomfortable for anyone hoping the software vendor absorbs some of it. You do. The registrant whose name sits at the top of that entry owns every word in it, whether they typed it, dictated it, or accepted it from an ambient scribe that was listening while they took an FS out of the drawer. There is no shared liability model here. There is no “AI-assisted” footnote that shifts 30% of the burden onto a company in Shoreditch.

Which means the ai generated clinical notes sign off step is the entire safety system. And in most practices running these tools today, it is a button.

What the click actually looks like right now

Here is a real pattern from practices running ambient documentation. The clinician finishes the appointment, walks to the next surgery, and at some point in the next hour opens a queue of six drafts. Each one is reviewed at roughly eleven seconds. Six notes, a bit over a minute, queue cleared.

Eleven seconds is enough to confirm that a note exists and that it reads like a dental note. It is not enough to confirm that the note describes what happened to that patient.

The failure mode is not the one people expect. Everyone braces for hallucination, for the model inventing a treatment that never occurred. That happens, but it is rare and it is loud: you notice “extraction UR6” in a note about a scale and polish. The dangerous errors are quiet and structurally plausible.

Tooth number transposition. The clinician says “lower left six” and the transcript captures “LL6”, but the charting integration writes to LR6 because of a mapping error between the scribe’s output and the practice management system’s quadrant convention. The note reads perfectly. The chart is wrong. Six months later a different clinician plans treatment off that chart.

Confident omission. The patient mentions, in passing, that they have started apixaban. It comes out mid-sentence while they are settling into the chair, before the formal history-taking, and the model treats it as small talk rather than clinical content. The note is complete, well-structured, and silent on a fact that changes your extraction plan. Nothing looks wrong, because nothing is there to look wrong.

Consent flattening. You spent four minutes on why you would not place an implant in that site, covering bone volume, smoking, and the maintenance burden. The note says: “Options discussed, risks and benefits explained, patient consented to treatment.” That sentence is worthless in a complaint. It is a template that happens to be true, which is exactly what a claimant’s solicitor loves to find.

Radiograph findings drifting. Tools in the radiograph-reading space (Pearl’s Second Opinion, Overjet, VideaHealth) output findings with confidence scores attached. When those findings flow into a note, the score often does not travel with them. “Interproximal radiolucency DB of UR5” is a different clinical statement to “interproximal radiolucency DB of UR5, 0.61 confidence, clinically unconfirmed”. The first is a diagnosis you have made. The second is a machine suggestion you have not yet evaluated. Once it is in the note under your name, it is the first one.

None of these get caught by reading fast. All of them get caught by reading in a specific order.

The under-a-minute read

The check below takes 40 to 55 seconds for a routine exam and restorative appointment, rising to about 90 seconds for a complex consultation or anything involving consent for an irreversible procedure. It works because it does not ask you to read the note. It asks you to check four specific things, and reading the note is a side effect.

Step one, roughly 10 seconds: the tooth numbers, out loud or under your breath. Not the whole note. Every FDI or Palmer notation in the entry, read against your memory of the appointment and against what is on the chart. If you restored two teeth, there should be exactly two teeth in the note, and they should be the two you restored. This single step catches transposition, which is the error most likely to cause downstream clinical harm and the one least likely to be noticed by skim-reading, because a wrong tooth number is still a valid tooth number.

Step two, roughly 15 seconds: the medical history delta. Ask yourself one question: did anything change about this patient’s health since the last visit, and is it in this note? Not “does the note have a medical history section”, which it always does. Did the patient say anything, at any point, about a new medication, a new diagnosis, a pregnancy, a hospital admission, a change in smoking or alcohol. If yes and it is not in the note, type it in now. The model missed it because it was said casually, and casual is how patients disclose most of the things that matter.

Step three, roughly 15 seconds: the consent paragraph, read as a stranger. Read only the sentences describing what you discussed with the patient, and ask whether someone who was not in the room could tell which patient this was. “Risks and benefits discussed” fails. “Discussed that the LL7 is heavily restored and the crown may fail within 5 years, patient wants to try it before considering extraction and an implant, understands the cost is not recoverable if it fails” passes. If it fails, write one sentence in your own words. One sentence. You are not rewriting the note, you are adding the specificity the model could not invent.

Step four, roughly 10 seconds: any AI-originated clinical finding gets a verb. If a radiograph tool contributed a finding, the note must say what you did with it. Confirmed, rejected, monitoring. Never leave a machine suggestion sitting in the note as a bare noun phrase.

BEFORE (accepted draft, 11-second review)
---------------------------------------------
BW radiographs taken. Caries detected UR5 distal,
LL6 mesial. Restorations sound. OH good.
Patient advised.

AFTER (40-second structured read)
---------------------------------------------
BW radiographs taken. AI screening (Second Opinion)
flagged UR5 D and LL6 M. On review: UR5 D confirmed
radiographically and clinically, into dentine, plan
composite. LL6 M flagged at low confidence, appears
to be cervical burnout on my read, NOT treating,
review at 6/12 recall with repeat BWs.
MH: patient now taking apixaban 5mg BD since Aug,
cardiology, added to MH. Relevant to any future XLA.
Discussed UR5 with patient, agreed to restore, booked
25 mins.

The second version took under a minute to produce from the first. It is also a note that defends you, charts correctly, and tells the next clinician something useful.

Why “review and approve” workflows fail in practice

Every ambient scribe ships with an approval step, and every vendor will tell you the clinician is in the loop. The loop is real. The problem is what the interface does to attention.

A queue of drafts presented as a list, each one green-ticked when approved, creates exactly the psychological conditions for rubber-stamping. It is a task with a visible completion state and no friction. Compare that to the old workflow, where you wrote the note and the act of writing forced you to reconstruct the appointment. The reconstruction was the safety mechanism, and it was free, because it was inseparable from the task. Removing the writing removed the reconstruction, and the approval click was supposed to replace it. It does not, because clicking is not reconstructing.

So the fix is not more diligence. Telling associates to be careful is not a control, it is a wish. The fix is to reintroduce a small amount of structured friction at a point where it costs under a minute.

Three things that make the check stick:

Sign off at the chair, not at the end of the session. A note reviewed while the patient is still walking to reception is reviewed against a live memory. A note reviewed at 6pm is reviewed against the note itself, which is circular and catches nothing. If your software allows a draft to sit in a queue for hours, that is a workflow problem you can fix without changing vendors.
Never approve more than one note at a time. If your system offers a bulk-approve or select-all, treat it as a defect and turn it off. Practices that have had bulk approval available have, in my experience of hearing about this, used it within about a fortnight.
Audit 10 notes a month, at random, across every clinician. Pull them, read them properly, compare against the chart. You are looking for one thing: whether the notes from different clinicians have started to sound identical. When they do, everyone is accepting the default phrasing, and you have a systemic documentation problem rather than an individual one.

That last point matters for the practice principal specifically. Your indemnity exposure is not one associate’s bad note. It is twelve months of notes that all say “options discussed, risks and benefits explained” because the model produces that sentence when the consent conversation was not captured well, and nobody pushed back. If you are building the wider system for how your practice records and charts treatment, the clinical notes and charting foundations are worth getting right before you layer AI on top, because AI accelerates whatever your documentation culture already is.

The DCP question, and the front desk

A separate problem, and one that practices get wrong more often than the clinical one. Reception-facing AI tools now draft notes too: triage summaries from phone calls, appointment reasons, symptom descriptions taken by a chatbot before the patient arrives.

These entries land in the patient record. Someone has to own them. The safe rule is that a note authored by an AI triage tool is unsigned patient-reported information until a registrant reads it and converts it into a clinical entry. It should be visibly marked as such in the record. A triage summary saying “patient reports severe pain UL, likely abscess” is not a diagnosis, and it must not sit in the record formatted like one, because in six months nobody will remember it came from a chatbot.

Practically: keep AI triage output in a separate, clearly labelled field. Do not let it merge into the clinical note. When the patient arrives, the clinician reads it, verifies it, and writes their own finding. The triage note stays as a record of what the patient said, which is genuinely useful, and never becomes a record of what you found.

What to ask a vendor before you buy

Four questions, and the answers tell you more than any demo:

Can a note be approved without the approval screen having been scrolled to the bottom? If yes, the product is designed for throughput, not safety.
Does the audit log record who approved each note, when, and how long the approval screen was open? The duration field is the one that matters. If it does not exist, you cannot demonstrate to anyone that review actually happened.
When an AI radiograph finding enters the note, does the confidence score and model version travel with it? If not, you cannot reconstruct why a finding was accepted or rejected.
Where is the audio, and for how long? Ambient scribes record consultations. Under UK GDPR that recording is special category data and you are the controller. Get the retention period in writing, get the processing location, and get the DPA signed before the first patient is recorded, not after.

If a vendor cannot answer question 2 with a screenshot of the log schema, they have not thought hard about the thing that will matter most when a case comes in.

Start Monday

Pick one clinician. Ask them to run the four-step read on every note for two weeks and keep a tally, on paper, of how many drafts needed a correction. The number is almost always higher than the practice expects, and it is the single most persuasive piece of evidence for getting the rest of the team to stop clicking through.