Ahpra AI guidelines: what patient data stays out of ChatGPT

Ahpra’s AI guidance doesn’t ban ChatGPT or AI scribes from a practice. It reads the codes practitioners already answer to against a new kind of software, and leaves every consequence with the person who signs the note.

That puts two very different tools under one set of rules: the scribe listening in the consult room, and the chatbot a tired clinician opens at 7 pm to tidy a letter. Which accounts hold up under the Privacy Act is in the guide to choosing an AI tool.

Disability support has its own version of the question, in what to take out of an NDIS case note.

What are the Ahpra guidelines for using AI in practice?

The page is called Meeting your professional obligations when using Artificial Intelligence in healthcare, reviewed 22 August 2024. It creates no offence. Through five principles, it shows how the existing codes of conduct apply to AI, and a complaint about a leaked note would be judged against the code.

Which code depends on your profession. Twelve National Boards share one Code of conduct, from physio and OT to pharmacy. Doctors, nurses, midwives and psychologists have their own board’s code, which is what Ahpra’s “or equivalent” covers. Under all of them, health details are sensitive information under the Privacy Act.

The principle people skim is understanding. Ahpra expects you to know where a tool keeps data and whether it retrains on it, before the first patient. That’s a vendor question nobody else in the practice is going to ask for you.

Is it illegal to put patient information through AI?

Not as such: no Australian statute names ChatGPT. But size is no escape. Section 6D(4)(b) of the Privacy Act says anyone who provides a health service and holds health information isn’t a small business operator, so a two-room physio clinic is covered like a private hospital group.

Health information is sensitive information under section 6(1). The OAIC’s guidance on commercially available AI, published 21 October 2024, says privacy obligations attach to anything put into an AI system. It recommends against entering personal, and particularly sensitive, information into publicly available generative AI tools.

So the paste is a use or disclosure that has to fit APP 6. A Medicare number adds APP 9, which allows a government related identifier out only for listed reasons, like verifying identity. Polishing a letter isn’t one. Whether it’s also a notifiable breach is a separate test.

Can ChatGPT be used for medical notes?

For the prose, yes. The RACGP’s conversational AI resource, first published June 2025 and updated September 2025, lists clinical letters, notes and discharge summaries among the jobs these tools can help with. The same page says GPs should never enter sensitive or identifying data into them.

Ahpra’s first case study is exactly this: findings dictated into ChatGPT. The data may sit offshore, and the patient would have to agree explicitly that others could use, view or sell it. Taking the patient out is easier than that chat. See what OpenAI keeps from a paste.

Names and numbers are the easy half. The story is harder: the supermarket, the month, the husband’s first name. Cutting those is a judgement call, not a legal finding, because de-identified has a stricter meaning. Here is an invented psychology note, line by line.

Labels alone leave text that is pseudonymised, not de-identified; the Australian difference is in de-identified vs pseudonymised.

Other numbers hide in referral packs too, like the card number on a NSW licence attached for identity.

If identifiable text already went in, deleting the chat clears your history, not the disclosure.

Do you need patient consent for an AI scribe?

Yes, every time. The RACGP’s AI scribes resource, last updated October 2025, says get consent at the beginning of each consultation. Ask your MDO whether it wants that in writing; if not, verbal consent recorded in the notes will do, and some scribes write that line themselves.

Ahpra explains the timing. Recording a private conversation without consent can carry criminal implications, so the software should ask for explicit consent as its first step. Ahpra also wants the answer noted in the health record, which makes a scribe’s own consent line useful, not decorative.

A waiting-room poster isn’t the yes. The OAIC says consent can’t be implied just because someone was notified: signage tells patients, the question in the room asks them. Accountants settle it up front with an AI clause in an engagement letter; a consult can’t.

Can you say no to your doctor using an AI scribe?

Yes. The OAIC’s guide to health privacy says valid consent is informed, voluntary, current, specific, and given with capacity. A yes the patient couldn’t refuse fails the voluntary part. If no means a shorter consult or a raised eyebrow, the practice has a design problem, not a difficult patient.

Victoria’s Surveillance Devices Act, s 6(1), catches recording a private conversation you’re not a party to without each party’s consent. The clinician is a party; whether the vendor’s software counts as someone else is a lawyer’s question. The RACGP says some jurisdictions make recording without consent an offence.

Ahpra and the RACGP say nothing about carers or interpreters. My rule is stricter: everyone who’ll speak gets asked, and one no keeps the scribe off. Support workers often sit in, which is why the NDIS providers page treats care teams as part of the record.

Is doctor-patient confidentiality a law in Australia?

It’s both a professional duty and, in parts, the law. Section 3.3 of the shared Code calls privacy an ethical and legal obligation. It asks for informed consent before disclosing, and written consent before sharing someone’s information even when they aren’t named.

The legal layer depends on who employs you. A private practice answers to the Privacy Act. The Act’s definition of organisation leaves out state authorities, so public hospitals sit under state privacy law, the split that decides which privacy law covers a council.

That split is the least intuitive part to me: the same note falls under a different Act depending on who pays your wages. For a question about one patient’s record, ring your MDO before acting on anything here.

An AI policy for a healthcare practice: scribes, chatbots and the TGA

The RACGP puts the scribe decision with the practice owner: write down which scribes are allowed, tell tenant GPs and staff, and say so if the answer is none. A scribe policy silent on ChatGPT is half a policy. A starting draft is in our AI policy template.

Ahpra and the RACGP both say scribes usually aren’t medical devices, because they summarise what was said. Ahpra’s second case study is the exception: a scribe added a plausible diagnosis to a referral letter. That met the definition, so the tool should have been on the ARTG unless an exemption applied.

An invented diagnosis is the feature I’d want switched off before rollout, and a practice-wide setup makes that one decision instead of twenty.

Keeping patient data out of ChatGPT before the letter leaves your screen

Most of this comes down to a habit at the keyboard: the header, the numbers and the giveaway details out, the clinical question in. Outside the consult room the list of identifiers runs longer, and the Australian ones a document tends to carry are worth knowing by sight.

A referral letter carries most of the patient in its header, so that’s the part worth automating. Nonimo for an Australian practice clears the name, Medicare number and contact lines from the draft before it reaches the chatbot. The shop, the town and the month are still yours to rewrite.

None of the draft is sent to us; the security page covers how. Paste a letter below and see which fixed-format numbers survived your own edit.

Sources

• Ahpra, Meeting your professional obligations when using AI in healthcare. Five principles; consent noted in the record; explicit consent step before recording; scribes usually not TGA-regulated. Page reviewed 22 August 2024.
• Ahpra, AI in healthcare case studies. ChatGPT dictation and offshore storage; a scribe adding a diagnosis meets the medical device definition.
• RACGP, Artificial intelligence (AI) scribes. Consent at the start of each consult, MDO and written consent, practice policy, privacy policy, overseas storage, TGA. Last updated October 2025.
• RACGP, Conversational artificial intelligence. Letters and notes as uses; never enter sensitive or identifying data.
• OAIC, Guidance on privacy and the use of commercially available AI products. Obligations attach to AI inputs; sensitive information out of public tools; consent not implied by notice.
• OAIC, Guide to health privacy: introduction and key concepts. Health information includes name and Medicare number; elements of valid consent.
• Privacy Act 1988, Federal Register of Legislation. Compilation No. 104: s 6(1), s 6D(4)(b), APP 6 and APP 9.2.
• Surveillance Devices Act 1999 (Vic). Section 6(1), version 048.

Accessed 7 November 2026.