When not to use
an AI scribe

Consent and clinical judgement answer different questions, and both have to be satisfied. Consent tells you the patient agrees to be recorded. Judgement tells you whether recording this consultation, with whoever is sitting beside the patient, changes what gets said or where the material could end up. A patient can give a perfectly valid consent to a recording that is still the wrong call, and in several of the situations below the patient is the last person able to weigh it.

The two possible errors do not cost the same, which is what sets the threshold. Leave capture off where it would have been fine and you lose a few minutes at the end of the day. Leave it on where it should have been off and you may suppress the disclosure the appointment existed for, or create material that reaches someone it never should have.

Most of the time the decision can be made before the patient sits down. The reason for the visit, who walked in with them, who booked the appointment, whether the last entry carries a safety flag, and whether this is a phone consultation are all visible in advance. Deciding then is far less conspicuous than deciding part way through.

The recognised technique for a consultation where a partner or family member is present and answering for the patient is to create a moment alone: the other person waits outside during the examination, or the patient comes to the treatment room for a specimen. The RACGP's guideline on abuse and violence in general practice, the White Book, sets that craft out in detail. The moment alone works because the patient believes what is said in it stays between the two of you.

Ambient capture cuts across that belief. A patient told the room is being recorded has been told that a file of the conversation exists somewhere, and someone living with a controlling partner spends a great deal of their life working out who can reach files. That caution is well founded: health records are read by patients, released under signed authority, and produced under subpoena in family law and criminal proceedings as a matter of routine.

The second problem is easier to miss. In a room with a suspected perpetrator, asking for consent to record is itself an act with consequences. The other person may answer for the patient, which tells you something and tells them more. The patient may agree because refusing would need explaining in the car afterwards. Do not raise it at all: keep capture off and document the appointment the ordinary way.

Phone and video consultations remove the check all of this rests on, which is seeing who else is present. On a phone call you cannot tell whether the handset is on speaker, and on video you see a frame rather than a room. Asking whether the patient is alone and free to speak is ordinary practice on those calls, and the answer bears on recording as much as on what you ask next. Three working rules keep this manageable in a busy list:

• If capture has already started and the picture changes, stop at the first natural break rather than waiting for the end of the consultation.
• Do not use the consent question to test the room. It is not neutral in front of a controlling partner.
• Document the safety concern through the process your practice already uses. Restricting the visibility of an entry in your clinical software does not keep it out of a record produced under subpoena.

Mandatory reporting is set by each state and territory, and the schemes differ in who is a mandated reporter and in which kinds of harm must be reported. Doctors and nurses are mandated reporters in every state and territory. Past that the lists diverge, with some naming psychologists, dentists or pharmacists and others not, and in the Northern Territory the duty sits on every person rather than on a list of occupations. Several jurisdictions also carry a separate criminal offence for an adult who fails to report a known child sexual offence. Check the obligation where you practise rather than where you trained.

A consultation where a child protection concern is forming is one where you are gathering information that may end up in a notification, in a court, and in front of a family reading every word closely. Children and young people disclose to a clinician they trust in a room they believe is private, and adolescents test that privacy carefully first. Telling a fifteen year old the conversation is being recorded changes the odds that the disclosure happens at all.

Write the notification yourself as well. It carries your clinical reasoning: what you observed, what was said and by whom, what you concluded, and why you formed the belief you formed. A summarising tool works by compressing spoken language into clinical shorthand, which strips out the specificity a notification is judged on. Draft it as close to the encounter as you can manage.

The triggers are visible before the consultation starts: an appointment booked by a school, a caseworker or under a court order; an injury whose history does not fit the injury; a young person brought in by an adult who wants to stay for the examination. An adolescent asking to be seen without a parent present is on the same list. Any one is reason enough to leave capture off before you have formed a view.

Patients rarely refuse outright. They hesitate: they say they suppose that is fine, they glance at the device on the desk, or they ask who will be able to hear it. Treat each of those as a decline, tell the patient it is off, switch it off, and carry on without making it a discussion. The alternative is a patient who spends their fifteen minutes managing your comfort instead of their own problem.

The harder version is the consultation that starts as one thing and becomes another. A patient books for a repeat prescription and twenty minutes in tells you about a sexual assault. Capture has been running the whole appointment by then. Stop it, then deal with what has already been taken: discard the session rather than leaving a partial recording in a queue for review.

There is also a documentation reason to write these notes by hand. A record that may be tendered in evidence needs the patient's own words, in the order the patient said them, alongside the questions you asked. Ambient tools paraphrase spoken language into clinical language, which is what makes them useful in an ordinary consultation and wrong for this one. A formal forensic medical examination is specialist work with its own chain of custody rules, and no scribe belongs in it.

One rule covers the whole group. Leave capture off whenever the patient's willingness to speak is the clinical value of the encounter. That takes in disclosure of assault, sexual and reproductive health, drug and alcohol use, gambling, immigration status, and anything the patient has not said out loud to anyone before.

For a patient with persecutory delusions, particularly delusions carrying a surveillance theme, being told that a device in the room is listening confirms the belief you are there to assess. You will either lose the therapeutic contact for that appointment or feed the delusion. The same reasoning covers a patient who is acutely agitated, intoxicated or delirious. Capacity is decision specific, and agreeing to be recorded is one of the decisions that cannot meaningfully be made in that state.

Risk assessment is the second reason to keep capture off in acute mental health work. A suicide risk assessment is a clinical formulation rather than a record of what was said, and the note has to show what you asked, how you weighed static and dynamic factors, what you concluded and what you did about it. These are among the notes most likely to be read years later by a coroner. A drafted summary that renders a careful, qualified assessment as a tidy line about the patient denying plan or intent is a real medico-legal exposure, and review does not reliably catch it, because the line reads well.

Assessments and orders made under a state or territory Mental Health Act sit outside this work entirely. They are statutory instruments completed by the examining practitioner, recording the grounds observed and the criteria applied, and signed as a personal attestation that you formed the view yourself. Fill them in yourself, every time, and do not paste from a draft.

None of this rules ambient documentation out of mental health practice. A stable review, a medication discussion and most therapy sessions where the client has agreed behave like ordinary consultations. The exclusions are the acute presentation and the risk assessment, and both are recognisable at the time.

Treating a colleague, a staff member, a staff member's child, or a patient everyone at the front desk recognises is a documentation problem before it is anything else. The Medical Board's code of conduct tells doctors to avoid providing care to anyone they have a close personal relationship with, and the other national boards take the same line on boundaries. Where care of that kind is unavoidable the record deserves more attention rather than less. More people have a plausible reason to open that file, and a scribe adds places the material sits: an unsigned draft in a review queue, and audio held for whatever retention period applies.

Settle one question before the consultation rather than during it: who can open an unsigned draft, and who can access or replay audio. Most practices have never asked. Permissions on the clinical record are usually well understood, permissions inside the scribe's own workspace usually are not, and they are separate systems with separate access lists. Until you can answer that in writing, document staff and colleague consultations the ordinary way.

Rural and remote practice compresses all of this. The receptionist is a neighbour, the practice nurse is the patient's cousin, and the town knows who was in the waiting room on Tuesday. Ambient documentation does not create that problem, but it should shift the default: any consultation touching a locally sensitive subject is documented the ordinary way unless the patient agrees without hesitation.

There is also the person in the room who is not the patient. Capture takes in whoever speaks, including a support person, an interpreter, and a parent who mentions their own diagnosis while their child is the patient. That material now sits inside somebody else's health record, and nobody asked them.

Know where the stop control is before you need it, and be able to reach it without looking away for more than a second. If stopping means unlocking a device, finding a tab and confirming a prompt, you will not do it in the moment that calls for it. Practise it on ordinary consultations so the habit is there on the day it matters. There is almost always a natural break to use: you stand to examine, or the patient steps behind the curtain. If you have to say something, keep it to one short sentence that puts the decision on you, such as telling the patient you will write this one up yourself.

Stopping and deleting are different actions. A stopped session may still hold everything captured up to that point, and if you stopped because the consultation turned sensitive then that partial capture is precisely what you do not want retained. End the session and discard it, and confirm in advance that your tool treats a discarded session as deleted rather than archived somewhere the practice cannot see. Confirm the same for the transcript, a separate artefact from both the audio and the note that can outlive both if nobody asks.

The note itself does not usually need to record that a scribe was not used. What the record has to show is the consultation: history, examination, reasoning, plan, and that it was written contemporaneously by the clinician who provided the care. The instrument used to produce the text is no more part of that than whether you typed the note or dictated it. Four things are worth handling deliberately:

• If you raised recording and the patient declined or hesitated, record the outcome where your practice keeps consent information, in neutral words such as declined ambient documentation. Do not record the reason the patient gave.
• If you stopped because of a safety concern, document that concern through your practice's usual family violence or child protection process. A line in the consultation note saying you switched the recording off because the patient seemed frightened of their partner is itself a disclosure, in a record that partner may one day read.
• If your practice reports on scribe usage, make sure a low rate is read as clinical judgement rather than poor adoption. Otherwise the metric quietly pushes clinicians to record consultations they should not.
• Name these categories in your practice's AI use policy, so that leaving capture off is a team standard rather than one clinician's habit, and so that a locum knows it on their first morning.