AI Medical Scribes in Australia: Accuracy, Consent and Regulation (2026)
By Russell Bewsell, Founder & CEO, Voice Recognition Australia • 28 years in clinical speech technology • Published 17 August 2026
In short: an AI scribe listens to your whole consultation and generates a note, so it can record things that were never said. Speech to text types only the words you deliberately speak, so it cannot invent clinical content. No AI scribe is TGA-approved, and AHPRA requires you to check every line either way.
ABC News reported on 14 August 2026 that a patient's medical record was altered by an AI scribe to state she had been micro-dosing psychedelic mushrooms. She had never taken them. She discovered it only when she read a post-operative letter her specialist had sent to her GP — and she was receiving workers' compensation at the time, so a fabricated reference to illegal drug use was not a trivial error.
That story has prompted a lot of questions from clinicians this week. This article sets out what is known, what the regulators currently require, and how ambient AI scribes differ from speech to text — because the distinction matters and is not well understood.
What is an AI medical scribe, and how is it different from speech to text?
An AI scribe listens to the whole consultation and generates a summary note. Speech to text software — also called dictation — transcribes only the words the clinician deliberately speaks. That difference determines what can go wrong.
An ambient scribe records the conversation between clinician and patient, then uses a language model to produce a structured note. The clinician did not compose that note — the model did, from an audio recording of a conversation. Speech to text takes the opposite approach: the clinician composes the note out loud, and the software types those words at the cursor.
| Ambient AI scribe | Speech to text (dictation) | |
|---|---|---|
| What it listens to | The whole consultation, including the patient | Only what the clinician chooses to say |
| Who composes the note | A language model, from the recording | The clinician |
| Typical error | Plausible-sounding content that was never said | A misheard word |
| Patient consent | Required — the patient is being recorded | Not applicable — the patient is not recorded |
| Admin time saved | Substantial — the note is drafted for you | Moderate — you still compose, but faster than typing |
| TGA approval | None approved as at August 2026 | Not a medical device — it is a typing method |
| Main risk | Fabrication, and over-trust once checking slips | Mis-recognition you must proofread |
What errors have actually been documented?
The documented AI scribe failures involve invented clinical facts, not just misheard words. Alongside the fabricated drug use in the ABC report, a Digital Rights Watch study cited a scribe recording the wrong breast in a breast cancer diagnosis, and recording epilepsy in a patient who did not have it.
The Consumers Health Forum has noted a behavioural risk that compounds this: clinicians check output carefully for the first few weeks, then stop checking once they assume it will be right. A Perth GP quoted by the ABC — himself an advocate of the technology — noted that a classic failure is the model reversing the side of the body, saying right when the clinician said left, and that drug doses need particularly close checking.
None of this means AI scribes are useless. The same reporting notes real benefits: less administrative burden, running to time more often, and more thorough referral letters. The point is narrower — the failure mode is different from what clinicians are used to, and harder to catch, because a fabricated sentence reads perfectly plausibly.
Are AI scribes approved by the TGA in Australia?
No. As at August 2026 no AI medical scribe has been approved by the Therapeutic Goods Administration. Products that claim only to transcribe and summarise fall outside the TGA's definition of a medical device, so they are exempt from that oversight. The TGA treats a tool as a medical device when it does things like suggest a diagnosis or treatment plan.
The TGA has acknowledged that some scribes claiming to offer clinical suggestions may therefore be operating in breach, has spent the past year reviewing the category, and has said findings will be released in coming weeks with breaches acted upon. Digital Rights Watch has called for the AI Safety Institute to test scribes and publish an approved list — the approach New Zealand has taken, where two tools were approved for clinician use.
What AHPRA expects of you right now
AHPRA's position is that clinicians must check all output from an AI scribe for accuracy in order to meet their professional obligations. The regulatory gap around the products does not transfer to you — responsibility for what ends up in the record remains the clinician's, regardless of which tool produced the draft.
Do I need patient consent to use an AI scribe?
Yes — and a sign in the waiting room does not meet the requirement. Digital Rights Watch found wide variation in how consent is being obtained, including notices in waiting rooms that do not satisfy privacy law, consent bundled into online booking, and — most concerning — patients being refused appointments if they declined.
The Consumers Health Forum position is that consent must be enthusiastic and fully informed, and that patients should feel able to ask questions. There is also a clinical-quality argument: the Digital Rights Watch report found patients may withhold sensitive disclosures, such as family violence, when they know AI is listening. A patient who does not tell you something is a worse outcome than a slower note.
Questions patients are being encouraged to ask
- Are you using AI today, and what for?
- Will you review everything before it goes into my record?
- Can I get a copy of the notes?
- What happens if I think something is wrong?
If your practice cannot answer those four cleanly, that is worth resolving before the next patient asks.
What is the alternative to an AI scribe?
The alternative is speech to text: you speak the note yourself and the software types exactly what you said, so nothing enters the record that you did not say. It is the same technology clinicians have long called dictation, but the current generation needs no voice training and works directly inside your EMR.
This removes the fabrication risk, but it does not remove your obligation to proofread. Speech to text can still mis-hear — a drug name, a dose, a laterality — and those errors are exactly why the note still has to be read back before saving.
That is a genuinely different risk profile, not an absence of risk. A misheard word is usually visible on screen as you speak. A fabricated sentence in a generated summary is not, because it reads like something you might plausibly have said.
The trade-off is honest: an ambient scribe saves you more time, because it drafts the note for you. Speech to text saves less, because you still compose it. What you get in exchange is that nothing enters the record that did not come out of your mouth.
Where Speech Recognition Cloud fits
Speech Recognition Cloud Medical is verbatim speech to text, not an ambient scribe. It does not record the patient, does not listen to the consultation, and does not generate summaries. Its AI writing features are deliberately excluded from the Medical tier, so no language model rewrites clinical text. It types at your cursor in Best Practice, Medical Director, Genie, Zedmed, Helix and any other Windows application.
Being straight about the limits: it is Windows 10 and 11 only, it needs an internet connection, and the medical vocabulary is English only.
The free 30-day Medical trial attaches to your email address — you install the standard build and it becomes the Medical edition. No separate download, no licence key, no credit card.
Should I use an AI scribe or speech to text?
- An ambient scribe may suit you if consultation volume is your main pressure, your consent process is genuinely robust, and you have the discipline to keep checking output after the novelty wears off.
- Speech to text may suit you better if your notes are detailed prose, if you work in areas where patients disclose sensitive material, or if you are not comfortable with a model composing clinical content on your behalf.
- Either way, you are accountable for the final record. AHPRA's expectation does not change based on which tool drafted it.
Frequently asked questions
Are AI medical scribes approved by the TGA?
No. As at August 2026, no AI medical scribe has been approved by the Therapeutic Goods Administration. Products claiming only to transcribe and summarise are exempt from medical device oversight. The TGA has been reviewing the category and has said findings will be published in coming weeks.
Can an AI scribe put something in my notes that was never said?
Yes. Documented cases include a fabricated reference to a patient micro-dosing psychedelic mushrooms, a scribe recording the wrong breast in a breast cancer diagnosis, and epilepsy recorded in a patient who did not have it. AHPRA requires clinicians to check all AI scribe output for accuracy.
Do I need patient consent for an AI scribe?
Yes. Consent must be informed and freely given. A sign in the waiting room does not meet privacy law requirements, and refusing appointments to patients who decline is not an acceptable practice.
What is the alternative to an AI scribe?
Speech to text software, historically called dictation. The clinician speaks the note and the software types those words at the cursor, so no content can be generated that the clinician did not say. It still has to be proofread, because speech to text can mis-recognise words including drug names and laterality.
Is speech to text safer than an AI scribe?
Speech to text cannot fabricate content because it types only the words the clinician deliberately speaks. It can still mis-recognise words, so the note must be proofread before saving. It is a different risk profile rather than no risk.
Does speech to text record the patient?
No. Speech to text transcribes only what the clinician speaks into their microphone. The patient is not recorded and the consultation conversation is not captured, so patient consent for recording does not arise.
What percentage of Australian GPs use AI scribes?
The Royal Australian College of General Practitioners estimated last year that around 40 per cent of GPs were regularly using AI scribes, and that figure was considered conservative at the time.
Sources and further reading
- ABC News — Error by AI scribe during medical appointment leaves patient devastated, 14 August 2026
- The Guardian — Doctors, AI scribes and the government privacy warning, 5 July 2026
- Digital Rights Watch — report on AI medical scribes in Australia
- Royal Australian College of General Practitioners — estimates on AI scribe adoption
- Therapeutic Goods Administration — medical device classification and AI scribe review
- AHPRA — professional standards on the use of AI in clinical practice
Try verbatim medical speech to text free for 30 days
Full Medical Ultra functionality. No credit card. Install the standard build — the entitlement attaches to your email.
This article describes the regulatory position as at 17 August 2026. The TGA review of AI scribes was pending at time of writing and this page will be updated when findings are published. Nothing here is legal or clinical advice — check your own organisation's policy before deploying any tool on a clinical device.
Russell Bewsell is Founder and CEO of Voice Recognition Australia and has worked in speech recognition since 1998, deploying it across Australian hospitals, GP practices, courts and government departments. He presented the national Dragonology clinical training series across Australia in 2017.