AI scribes in Australian consulting rooms: consent, notes and medico-legal risk
Ambient AI scribes are arriving in Australian consulting rooms faster than the consent scripts. This article covers AHPRA and privacy issues, who authors the note, hallucination risk, and a practical way to ask patients before the microphone starts listening.
Dr Elise Tan, GP and digital health editorUpdated

The pitch is easy to like. A microphone in the room, a draft note waiting when the patient leaves, more eye contact, fewer clicks, home in time for dinner. Ambient AI scribes — tools that listen to a consultation and produce a clinical note — are spreading through Australian general practice and specialist rooms for the most human reason in medicine: documentation is crushing people.
The risk is just as easy to under-describe. The tool is sitting in a conversation that includes mental health, sexual history, family violence, a teenager who asked the parent to wait outside, and a passing mention of a relative's cancer. It is sending audio or a transcript to a vendor. It is drafting a document that may later be read by a court, a health complaints commissioner, a colleague, or the patient. And it can, with a straight face, write down a symptom that was never said.
This is a piece about consent, authorship and the unglamorous review step. It is not a buyer's guide and not a prohibition. Some clinicians will use these tools well. Some are already using them as if a fluent paragraph were the same thing as a true one.
What an ambient scribe actually does
The product category is broader than the marketing word. Some tools record audio and send it away for transcription and summarising. Some process on a device. Some listen continuously. Some start only when you press a button. Some integrate with practice software. Some live in a browser tab and ask you to paste.
All of them sit in a consultation that used to be two humans and a keyboard. The keyboard was slow. It was also obviously a thing the clinician was doing. A microphone is quieter. Patients do not always notice quiet things, which is precisely why you have to say them out loud.
The output is usually a structured note: history, examination, impression, plan. It may propose a letter to a specialist, a patient summary, or a list of actions. Fluency is the feature and the trap. A messy consultation can become a tidy story. Tidy stories are how omissions hide.
If your cybersecurity floor is still a shared password and an untested backup, adding a scribe is not innovation. It is another vendor with a copy of the most sensitive hour of the patient's week.
Privacy, AHPRA and the law you already had
Health information is sensitive personal information under the Australian Privacy Principles. Collecting it has to be reasonably necessary, fair, and — in ordinary clinical practice — expected by the person in front of you. Sending that information to a vendor, possibly outside Australia, is a further use and a further disclosure. APP 5 notices and APP 8 overseas disclosure are not optional flavour text because the tool is fashionable.
The Office of the Australian Information Commissioner does not need a new statute to take an interest in a clinic that recorded consultations without telling people, or that could not say where the audio went. The My Health Records rules are not the main character here. The ordinary practice file is. Most scribe output will live in local clinical software, not in the national record.
AHPRA's public messages about AI have been consistent in spirit even as the products change: you remain responsible for your practice. You cannot outsource clinical judgment to a model. You cannot pretend a generated note is "just a draft" if you have signed it and billed against the encounter. Professional standards about records — contemporaneous, accurate, respectful — still apply when the sentences arrived in a burst.
Recording laws sit underneath, and they vary by state and territory. In some places, recording a private conversation without the consent of the parties is a criminal problem, not only a privacy one. A clinic that relies on a waiting-room poster and a microphone that is always on is taking a legal bet. Get local advice. Do not take it from a vendor slide.
Medicare is not a privacy regulator, but it is a billing system that expects the record to support the service you claimed. A hallucinated examination is not only a clinical lie. It is a documentation lie. If you did not look in the ear, the note should not say you did, however confident the prose.
Consent that is more than a sign
A sign on the wall that says "we may use digital tools" is the bare minimum and, used alone, a bit of a dodge. People do not read signs when they are waiting to talk about a breast lump. Consent for a scribe should be spoken, specific, and easy to refuse without punishment.
Specific means: a microphone will listen; a company will process the recording or transcript; the purpose is to draft my notes; you can say no and I will type; this is not being posted on the internet, and it is also not staying only in my head.
Easy to refuse means the GP does not sigh, joke, or imply that saying no will shorten the appointment. If the business model of the clinic depends on every patient agreeing, the business model is the ethical problem.
Some consultations should default to off. A first disclosure of assault. A child or adolescent where confidentiality is already fragile. A person who is paranoid, for whom a hidden microphone is not a metaphor. A colleague sitting in for a teaching session who did not agree to be in a vendor's dataset. A patient who asks, simply, not to be recorded.
Carers and interpreters complicate the room. If a third adult is speaking, they are in the recording too. Ask them. If a health app on the patient's phone is also listening — a symptom diary, a translator, a relative on speaker — you now have several processors and no single consent story. Keep the room simple.
A practical script that does not sound like a lawyer
You do not need theatre. You need a pause before the clinical work starts.
"I use a note-taking tool that listens to our conversation and drafts the record so I can look at you instead of the keyboard. A company processes the audio to do that. I read and correct the note before it stays on your file. You can say no and I will type as usual. It will not change your care. Would you like me to use it today?"
If they hesitate, turn it off. Hesitation is information.
If they ask where the data goes, answer honestly or do not use the tool. "I'm not sure" is not a consent conversation. It is a reason to email the vendor and stop until you know whether processing is onshore, whether recordings are retained, and whether the company trains models on clinical speech.
If they ask whether the recording can be played later in a dispute, do not invent comfort. Some vendors retain audio. Some retain transcripts. Some claim not to. Your contract and settings decide. You should know before the microphone is on, not after a subpoena.
Hallucination is not a quirky side effect
Large language models generate plausible language. That is the job. Plausible is not the same as true. In a clinical note, the failure mode is ugly: a symptom added, a negative examination invented, an allergy omitted, a plan that sounds like a guideline and was never discussed, a family history borrowed from a different pattern.
Clinicians already know that dictated notes can be wrong. Voice-to-text has been turning "hypertension" into entertainment for years. Ambient scribes add a narrative layer. They do not only mishear. They summarise, infer, and tidy. Inference is where a mention of "my father died young" becomes a documented familial cardiomyopathy you never diagnosed.
The defence "I was busy, the AI wrote it" will not travel. The record is yours. Review it as if a clever medical student wrote it after sitting in — helpful, fluent, not to be trusted on the details that matter.
Particular danger zones:
- Medicines, doses, and allergies
- What you actually examined
- What you actually promised to do
- Mental health and risk statements
- Anything that attributes blame or describes a third party
- Letters that will leave the building
If you use the scribe to draft a specialist letter, read the letter. Patients will see it. Specialists will notice the examination you did not perform. Complaints often begin with a document that surprised the person it describes.
Who authors the record
Australian clinical records are not a joint paper with a software company. The treating practitioner authors the note. The practice owns the record-keeping system. The vendor processes data. Those roles get mixed in conversation — "the scribe said" — and they must not get mixed in a complaint.
If you copy the draft in without reading, you have still authored it. If a registrar uses your account, you have a supervision and access problem as well as an AI problem. If the patient later asks for their record and finds a scene they do not recognise, the explanation "the model usually gets it right" is not an explanation.
There is a second authorship question: who is in the room on the recording. Students, practice nurses, family members, a child playing on the floor. Some of those people can consent. Some cannot. Some should not be identifiable in a vendor log at all.
Do not put scribe output straight into My Health Record or a shared letter as if the fluency proved quality. The national record and secure messaging multiply a mistake. For the rest of the digital-health context, including how Medicare telehealth after the pandemic already strained documentation, the pattern is the same. New channel, old duty.
A floor for clinics that will use these tools anyway
Most practices will not wait for a perfect national standard. Fine. Use a floor.
Choose a vendor that will answer, in writing, where processing happens, how long audio and transcripts last, whether they train on your data, how a patient access request works, and how you delete a consultation that should not have been captured. If they talk only about "enterprise-grade encryption", keep asking.
Turn the microphone off by default and on by choice, not the reverse. Give every clinician a way to pause when the conversation turns. Do not run the tool in the corridor.
Write the consent script into induction. Audit a sample of notes — not for style, for invented content and missing risk. Include scribes in your privacy policy in human language. Include them in your cybersecurity vendor register. Include them in your Notifiable Data Breaches thinking: a leaked inbox of generated notes is still a leaked inbox of health information.
Tell patients how to complain, including about the tool. A person who did not want to be recorded is not being difficult. They are doing your governance for you.
If you would not let an unseen intern file a letter under your name without a read-through, do not grant that privilege to a model because it sounds like one of the better registrars.
The point of the technology, if it has one, is to give attention back to the patient. That only works if the attention includes a minute of honest consent at the start and a minute of sceptical reading at the end. Everything between those minutes is a draft. Drafts are allowed. Unreviewed drafts that become the medical record are how a labour-saving device becomes a career-defining document.
Frequently asked questions
Who is the author of a note drafted by an AI scribe?
- The clinician. If you sign it, send it, or leave it in the record, it is your note. The vendor is not the treating practitioner and will not attend the coroner with you. Review the draft as if a tired locum wrote it on your letterhead, because that is the legal position in substance.
Do I have to tell the patient an AI scribe is listening?
- You should, and in many situations you must, if you want the collection of sensitive health information to be fair and expected. Recording laws also vary by state and territory. A posted sign is a start. A spoken question, with a genuine chance to say no, is the practice that will survive scrutiny.
What if the scribe invents a symptom I never heard?
- Treat that as a known failure mode, not a freak accident. Large language models can add plausible sentences. If you do not catch the extra chest pain or the missing allergy, the record becomes false. False records harm the next clinician and harm you when the note is read later in a complaint.
Corrections and complaints: read our editorial standards or write to the editors.