Ahpra expects a medical practitioner who uses an AI scribe to stay fully responsible for the record: inform patients, get informed consent before recording, understand how the tool stores and reuses data, check every note for accuracy, and confirm professional indemnity cover. Its guidance page is dated 22 August 2024 and will be updated, so read the live page.
The guidance does not ban scribes and does not approve any product. It says that existing professional obligations in your Board's code of conduct apply to new tools. That is good news, because most of the checklist below is things a careful practitioner already does, applied to a tool that writes in your name.
What the guidance actually says, in five principles
The regulator organises its guidance around key principles. Here they are in plain words, with the scribe version of each.
| Principle | What Ahpra says | What it means for a scribe |
|---|---|---|
| Accountability | The practitioner remains responsible for safe, quality care and must apply human judgement to any AI output. TGA approval does not change that. Tools should be tested for fitness before clinical use. | You are responsible for the accuracy and relevance of every record the scribe creates. Test it before you rely on it. |
| Understanding | Know enough about the tool to use it safely: its product information, how it was trained and tested, intended use, limitations, where data is stored and how it is used to retrain the AI. | Read the vendor's documents and be able to say where recordings and transcripts go and whether they train the model. |
| Transparency | Inform patients about your use of AI. Recording a consultation needs more explanation than, say, AI that helps interpret an image. | Tell patients what the scribe does, in words they follow, before the consultation starts. |
| Informed consent | Involve patients in the decision where their personal data is entered. Obtain informed consent, ideally noted in the record. Recording tools should include an explicit consent step first, and there may be criminal implications if consent is not obtained before recording. | No recording without a clear yes from the patient, and a note of that yes. |
| Ethical and legal issues | Confidentiality and privacy; whether patient data trains the model; employer or practice governance; relevant legislation; holding appropriate professional indemnity insurance and consulting your insurer if unsure about AI cover. | Check the data flow, your practice's governance, and your insurance, in writing. |
Ahpra's case studies sharpen the point. In one, a practitioner dictates clinical findings into a general-purpose generative AI tool and pastes the polished result into the notes; the guidance raises privacy, consent and accuracy concerns, noting that tools such as ChatGPT may store data outside Australia. In another, a scribing tool adds a plausible diagnosis the practitioner never made. Ahpra says that because that tool suggested a diagnosis, it meets the definition of a medical device and should be regulated by the TGA, which means inclusion on the Australian Register of Therapeutic Goods before supply unless an exemption applies. It adds that scribes can also produce invented content ("hallucinations") and omissions. By contrast, its main page says generative AI scribes are usually general-purpose tools that the TGA does not regulate, so for most scribes there is no register entry and the checking falls entirely on you. The practitioner remains responsible for the record. Everything in the checklist follows from those two cases.
The checklist, in the order to do it
Use the list as a gate. Do not move to the next group until the previous one is complete, and keep the evidence for each item in one pilot folder. If you work in a practice with a manager or a governance lead, share the folder with them.
Group 1: before you choose a tool
- Decide the purpose. Write one sentence: "The scribe will draft consultation notes for me to edit and approve." If you also want referral letters or coding suggestions, list them separately. Wider use means wider risk. Verify: the sentence is in the pilot folder and the vendor's intended use covers it.
- Read the vendor's product information. Look for intended use, limits, how it was tested, and contexts where it should not be used. Verify: save the documents with the date you read them.
- Find where data goes. Where are audio, transcripts and notes stored, in which country, for how long, who at the vendor can see them, and are they used to train or improve models? Verify: written answers from the vendor, not a sales call.
- Check regulatory status. Ask the vendor whether the product is a medical device and whether it is on the Australian Register of Therapeutic Goods. Ahpra points to the register as the place to check. If the tool suggests diagnoses or treatment, treat that as a flag. Verify: the vendor's written answer, and the register entry if there is one.
- Confirm the privacy position. Ask how the tool meets privacy and health records law for your state or territory, and what happens on a data breach. Verify: read the privacy terms yourself or have your practice's adviser do so.

The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowGroup 2: before you use it on a patient
- Confirm indemnity cover. Ask your insurer, in writing, whether use of this kind of tool in consultations is covered and whether they have conditions. Ahpra's guidance names holding appropriate professional indemnity insurance and consulting your provider if unsure. Verify: the reply is in the pilot folder.
- Check practice governance. If you work in a practice, hospital or group, confirm the policy allows it, who approves new tools, and your own role. Verify: an approval or a note that none is needed.
- Test it before clinical use. Run mock consultations with colleagues or staff playing patients. Include accents, interruptions, a medication list, a sensitive topic, and a patient who changes their mind mid-sentence. Verify: a short test log with errors found.
- Write the consent wording and the explanation. See below. Verify: a colleague outside the clinical team reads it and can say back what it means.
- Decide how consent is recorded. A tick box in the practice software, a line in the note, or a form. Verify: test it on three mock patients.
- Prepare an alternative. Patients who say no must receive the same care without it, with no awkwardness. Verify: your usual note-writing process is ready and the room is set up for it.
Group 3: in every consultation
- Explain, then ask, then record. Say what the tool does and where the recording goes, ask for consent, wait for a clear answer, and only then start the tool. Ahpra warns of criminal implications if you record before consent, so the order matters. Verify: the tool does not start until you press the button.
- Note the response. Ahpra says to ideally note the patient's response in the health record. Verify: a consent line appears in the record for each consultation where the tool is used.
- Stop at the patient's request. If a patient asks to pause or stop mid-consultation, you can. Verify: you know where the pause control is.
- Check every record before it is filed. Read the whole note, not a skim. The regulator holds you responsible for checking accuracy and relevance. Verify: a sign-off in your usual way, and the time you took.
Group 4: while it runs
- Keep an error log. Each error, its type (invented content, omission, wrong drug or dose, wrong side, wrong patient detail, an extra diagnosis), how you caught it, and what you changed. Verify: reviewed monthly.
- Re-check the vendor. New versions, new data terms, new sub-processors, and any change in where data is stored can change your answers to Group 1. Verify: a quarterly check with a date.
- Re-read Ahpra's page. The guidance says it will be updated regularly. Verify: note the date you last read it.
Consent wording that does the job
Consent has to be informed, which means the patient understands what is happening to their information and that they can say no. Short, plain and specific works better than a long form. Illustrative spoken wording, to adapt for your practice, your tool and your insurer:
"I use a computer tool that listens to our conversation and writes a draft of my notes. It means I can look at you rather than the screen. The recording and the draft notes are stored by [vendor] [for X days / in the practice system], and are [not] used to train their software. I check and correct every note myself. You can say no, and it won't change your care. Are you happy for me to use it today?"
The square brackets are the parts you must fill from the vendor's documents, and you must be able to defend each one. Never say "it's completely private" or "nothing leaves the practice" unless you have written confirmation that it is true. If you don't know whether recordings train the model, the honest wording is the one that matches the vendor's documents, and if you cannot get an answer, that tells you something about the vendor.
For a written version, add three lines: who can access the recording and notes, how long they are kept, and how to withdraw consent. For patients who need an interpreter, a carer or an easy-read version, adapt the explanation and record who was present. If a patient is a child or lacks capacity, consent comes through the appropriate person, as in your other consent processes. The tutorial on rolling out a scribe without losing trust has more on tone.
Checking the record: what errors look like
The regulator's case study names hallucinations and omissions. In practice, errors fall into recognisable types, and your checking should look for each.
| Error type | Illustrative example | How to spot it |
|---|---|---|
| Invented content | Note says "no known allergies"; nobody discussed allergies | For every negative finding in the note, ask whether you said it or the patient did |
| Omission | A symptom the patient raised at the end of the visit is missing | Re-read the last two minutes of your memory of the consultation; keep your own prompts, such as a list of the problems raised |
| Wrong number | Dose recorded as 50 mg; you said 5 mg | Check every dose, date, side and measurement against the prescription and your memory |
| Added diagnosis or plan | "Likely migraine" appears and you never said it | Scan the assessment and plan for words you did not use; delete them |
| Mixed-up speakers | The patient's history is attributed to a relative in the room | Check who said what when someone else is present |
| Template drift | Last week's phrasing repeated | Compare with the previous note for the same patient |
A realistic example: a practitioner dictates a plan to start a medicine at a stated dose, and the draft shows the right drug and the wrong frequency. The checking habit that catches it is reading every medicine line aloud against the script, every time. That takes ten seconds and is worth more than any feature on the vendor's page. The dictation or ambient notes tutorial compares the two styles of tool for a doctor's workflow, since each has different error patterns.
When a patient asks a hard question
Informed consent only works if you can answer follow-up questions. Prepare short, honest answers to the five that come up most, and keep them on a card beside the screen.
- "Who can hear the recording?" Answer from the vendor's documents: which staff, if any, can access audio, and whether it is deleted once the note is created. If you do not know, say so and offer to find out, rather than guess.
- "Is it used to train the AI?" Say yes or no according to the written terms. Ahpra's guidance specifically points to whether patient data is used to train the model and whether identifiable data finds its way into a learning database.
- "Where is it stored?" Name the country or say you are confirming it. Ahpra's case study flags offshore storage as a privacy risk.
- "What if it gets something wrong?" Explain that you read and correct every note, and that the patient can ask to see or correct their record in the usual way.
- "What if I say no?" Say that nothing changes about your care and you will take notes yourself.
Illustrative exchange: a patient says, "I'd rather it wasn't recorded, I'm here about something personal." The right reply is "Of course, I'll write my own notes", said warmly, with no second ask. Record the refusal in the notes in your usual wording. A practitioner who pushes a second time has turned consent into pressure, which defeats the purpose.
In a group practice: who signs what
If you work with colleagues, the personal checklist is not enough, because a scribe used by one practitioner affects the practice's data, its staff and its reputation. Agree three things in writing before more than one person uses it. First, who approves tools and keeps the vendor file: usually the practice manager or a named principal. Second, the standard consent wording and record, so patients get the same explanation from every practitioner. Third, the rule on what practitioners may not do, such as using personal accounts, pasting consultation content into free chat tools, or recording without the tool's consent step.
Illustrative: a four-practitioner practice puts the vendor file, the insurer replies and the pilot log in a shared folder with one named owner, and each practitioner adds their own error log. At the quarterly meeting, the owner reads the error logs together and asks one question: did any error type appear for more than one practitioner? A shared pattern points to a tool setting or template problem, and a one-person pattern points to a habit. Either is fixable, and neither is visible if every practitioner keeps their own counsel.
A four-week pilot plan
A pilot makes the checklist concrete and gives you evidence that you tested before relying on the tool.
- Week 0 (set-up). Complete Groups 1 and 2. Choose one practitioner and one consultation type, such as routine reviews, and avoid complex or sensitive consultations in the first fortnight.
- Week 1. Mock consultations only: five to ten, with a colleague as patient, and an error log. Fix settings and templates.
- Week 2. Live use with consenting patients, up to ten consultations. Check every record line by line. Time the check.
- Week 3. Extend to 20 consultations. Add the consent record check: for each consultation, confirm there is a consent line.
- Week 4 (review). Compare time spent on notes with the old way, count errors by type, count refusals, and gather any patient comments. Decide: stop, adjust or extend.
Set stop rules in advance. For example: any wrong medicine, dose or patient detail that reached the record unnoticed for more than a day, any consultation recorded without a documented consent, or any vendor change you cannot get answers about. Any one of these pauses the pilot until reviewed. The vendor security questions tutorial lists what to ask the vendor, and the is it worth it tutorial helps judge the time saved.
Worked example: a solo practitioner's evidence folder
Illustrative: a solo practitioner sees about 22 patients a day, spends around 5 minutes per note, and wants to cut after-hours note writing. She works through the checklist.
The folder after week 0 holds: the purpose sentence; the vendor's product information and data-handling answers (written, dated); a note that the vendor says the product does not suggest diagnoses and is not marketed as a medical device, with her own plan to check the register; the insurer's reply that use of a scribe in consultations is covered provided the practitioner reviews every note; and the consent wording and a consent tick box in her practice software.
In week 2, she records 18 consultations. Two patients say no. Her checking time averages 90 seconds per note, against 5 minutes of writing before, so on 16 notes the saving is about 56 minutes. She logs four errors: a missing symptom, a wrong frequency on a prescription, an invented "no allergies" line and a swapped speaker. All were caught on review, none reached the record. The week-4 review does the sum for full use: at about 100 consenting patients a week, saving 3.5 minutes a note comes to roughly 5.8 hours a week, mostly after hours, provided the 90-second check stays. All of these figures are illustrative. Measure your own, and don't assume the time saving holds if the checking gets sloppy. The saving only counts if the note is right.
What this checklist does not cover
It is a summary of one regulator's guidance, not legal or insurance advice, and it will age. Your state or territory privacy and health records laws, your employer's policies, your professional college's guidance and your insurer's terms may add requirements. Where they differ, follow the stricter one. If you work across several jurisdictions, or in a hospital, the governance steps may be more involved. And if your scribe starts doing more, such as suggesting diagnoses, writing referrals or coding, re-run the checklist for each new function, since the regulatory position can change when the tool's purpose does.
If you'd like help turning this into a plan for your own practice, that is what my AI implementation consultation covers: the checklist is the same, the evidence is yours.
Questions practitioners ask about Ahpra and AI scribes
Do I have to tell every patient I use an AI scribe?
Ahpra's guidance says practitioners should inform patients about their use of AI, and that a tool recording consultations needs more explanation than one working in the background. It also expects informed consent where patient data goes into a generative AI scribe, ideally noted in the record.
Is a TGA-registered scribe safer to use?
Ahpra says TGA approval does not change your responsibility to apply human judgement and check outputs. Most general scribing tools are not regulated as medical devices, but one that suggests diagnoses may be. Ask the vendor, or search the Australian Register of Therapeutic Goods, and still check every record.
Does my indemnity insurer need to know?
Ahpra's guidance asks practitioners to hold appropriate professional indemnity insurance for all aspects of practice and to consult their provider if unsure whether AI tools are covered. Ask your insurer in writing before use and keep the reply with your pilot records.
Can I use a free general chatbot as a scribe?
Ahpra warns that data entered into public generative AI may not stay private, and may be stored offshore or used by the developer. Free consumer tools are a poor fit for consultation content. Choose a tool whose data terms you have read and can explain to a patient.
Further reads
- AI Platform for Doctors: One Suite or Separate Scribe and Phone Tools? — One suite or separate scribe and phone tools.
- AI Scribe for GPs: Costs, Free Tiers and the NHS Pilot Limit — Costs and free tiers of scribes for GPs.
- AI Scribe Costs Compared: What Small Clinics Pay in 2026 — What small clinics pay for scribes.
- AI Scribe for Dentists: Will It Write Back to Your Software? — Whether a scribe writes back to your practice software.
- Telling Patients You Use AI Note-Taking: Consent Wording That Works — More consent wording that patients accept.
- AI Notes for Chiropractors: Consent and Data Checks Before Recording — Consent and data checks from another profession.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Ahpra, Meeting your professional obligations when using Artificial Intelligence in healthcare, and its case studies page (read 4 November 2026; page reviewed 22 August 2024). Always read the live page.