Issue 8 · 25 August 2026 · Fortnightly | AI in Healthcare · Australia & Beyond
This fortnight is mostly about your exposure. A patient discovered her medical record contained a drug-use claim she never made, and she found it in a letter already sent to her GP. Underneath that sits a number worth carrying into every vendor conversation you have this year: of 1,357 AI devices cleared for patient care, three have been tested on whether patients actually end up better off. The brighter news is where the automation is finally pointing — at the approvals layer, which is where your unpaid hours really go.
Here's what you need to know:
Summary
Read time: ~4 min
Top Stories
- An AI scribe fabricated a drug-use claim, and it reached a second clinician before anyone noticed
- Four health systems go live with real-time funding approval checks inside the record
- Only three of 1,357 cleared AI devices have been tested on whether patients do better
- AI applied to standard ECGs identifies a commonly missed form of heart failure
Signals
- The TGA moves from educating AI scribe suppliers to acting against them
- The NDIS Amendment Act commences this Thursday
- $430 million into Australian research, with AI among the named themes
- A regulator approves not just a model, but how it is allowed to change
Top Stories
Her Record Said She Was Micro-Dosing Mushrooms. She Found Out From the Letter to Her GP.
🇦🇺 Australia | Documentation Risk
ABC News reported the case on 14 August. The detail that should stop you is not the fabrication itself — it is where she found it. Green never saw the consultation note. She read a post-operative letter her specialist had written from that note and sent to her GP.
By the time anyone noticed, the invented claim had already left the originating clinician's system and been absorbed into another practitioner's record of her. Her specialist wrote that she takes documentation accuracy seriously but could not determine how the reference got in, attributing it to an error in dictation or transcription.
AHPRA's position leaves no ambiguity about who carries this. Using AI does not change your existing obligations — you remain responsible and accountable for your decisions and your records, you are expected to understand your tool's limitations, and inadequate verification of AI output can itself form the basis of a notification.
What it means for clinicians:
- Read the output against your memory of the session, not for whether it reads well. Hallucinated content is plausible by construction — that is the whole failure mode, and it is exactly why a quick skim slides straight past it.
- Extend the same check to everything you generate from the note: NDIS progress reports, referral letters, discharge summaries, funding submissions. An invention that survives into a second document has become someone else's record of your client.
- Say what the tool does when you take consent, not just that you are using one. Green agreed to transcription. Nobody asked her to agree to interpretation, and that gap is where this happened.
The Automation Worth Wanting Isn't Pointed at Your Notes
🌏 Global | Practice Admin
The mechanism pulls each insurer's authorisation requirements into the clinical record at the point of ordering or scheduling, so the clinician sees whether approval is needed before the appointment happens rather than after the claim bounces.
The AI is aimed at the genuinely miserable part: interpreting what the funder sends back, and reading the client's record to work out whether the requirements have already been met. The stated target is eliminating the phone calls, the faxes and the re-keying into separate portals.
Hold that against your own week. The industry has spent two years automating the clinical note, which was never where most of your unpaid time went. It went to establishing whether a support is claimable, chasing what a plan actually covers, and rewriting submissions that came back short.
What it means for clinicians:
- Ask any practice software vendor courting you what they are building for the approvals layer specifically. A scribe that writes a beautiful note does nothing for the hours you lose working out whether the service is fundable.
- You can capture some of this without new software. Moving the funding check to the point of booking, rather than the point of claiming, removes the rework — a template and a checklist get you a surprising share of the benefit.
- Watch which vendor brings this to NDIS and DVA first. The provider who can tell a client what is covered while they are still in the room has a real advantage, and it is a far better use of AI than another note generator.
3 of 1,357: What "Cleared" Actually Tells You About an AI Tool
🌏 Global | Vendor Claims
Clearance generally turns on whether a device performs as described and is broadly equivalent to something already on the market. It does not require anyone to show that using it changes what happens to the person in front of you, and for the overwhelming majority of these products, nobody has.
This does not mean the devices do not work. Accuracy against a reference standard is a real and necessary measurement. It is simply a different question from whether your client ends up better off, and the space between those two questions is where a great deal of clinical AI marketing comfortably lives.
The practical value here is a translation. "Cleared", "approved" and "included in the ARTG" all mean *performs as described*. None of them mean *shown to help*.
What it means for clinicians:
- When a vendor quotes accuracy, ask what changed for patients or clinicians once it was deployed. A specific answer is a genuinely good sign. A quick pivot back to accuracy figures tells you the outcome study was never run.
- Hold your own tools to the standard you would apply to any intervention you offer. If you cannot name what improved since you adopted it — minutes per note, rejection rate, hours on a Saturday — you are running on the feeling that it is faster.
- Treat the absence of outcome evidence as a reason to measure it yourself, not a reason to avoid the tool. A fortnight of before-and-after timings across your own caseload beats any vendor's brochure.
AI Is Finding Heart Failure in Tests Your Clients Have Already Had
🌏 Global | Diagnostics
The appeal is entirely in the input. Most clinical AI asks a health service to buy something, install something, and retrain people around it. This asks for none of that, because the ECG has already happened and the data is already sitting in the file.
The type of heart failure at issue is the one that hides. It is missed in routine care precisely because the standard read looks unremarkable, which is exactly the sort of pattern a model trained across very large numbers of recordings is well suited to catch.
For allied health this matters at the referral boundary rather than the point of care. Exercise physiologists, physiotherapists and cardiac rehabilitation clinicians all work with people whose cardiac status was characterised somewhere upstream — and upstream characterisation is what is getting sharper.
What it means for clinicians:
- Expect better-characterised clients rather than new equipment in your own room. The near-term effect of this class of tool on allied health is that people arrive with conditions already identified, not that you start reading ECGs.
- If you work in cardiac or pulmonary rehabilitation, it is worth asking referrers whether their service uses this kind of analysis. It changes who gets referred to you and how early.
- Note the pattern for judging future tools: the ones that extract more from data you already collect will reach practice far sooner than the ones requiring new hardware and new workflows.
Signals
🇦🇺 Australia | AI Regulation — The TGA will begin compliance action against suppliers of some AI clinical scribes after a year-long review found scope creep, inadequate monitoring and a lack of transparency, announced 3 August. No AI scribe currently sits on the ARTG and an estimated 40% of Australian GPs use one. The action targets suppliers, not you — but your own exposure runs through AHPRA, so ask your vendor in writing whether their product does anything the TGA would call a medical device. Read →
🇦🇺 Australia | NDIS — The Securing the NDIS for Future Generations Act was signed on 20 August and commences 27 August — this Thursday. Access rules are unchanged until 1 January 2028, so the near-term changes are to how plans are managed and how funding may be used. Read the funding-use rules against your current claiming practice this week, because clearer rules resolve ambiguity in both directions. Read →
🇦🇺 Australia | Research Investment — The NHMRC committed $430 million across 234 projects this month, funding researchers over five-year terms rather than single projects — long enough to carry a model from development to clinical validation. Ageing and chronic disease projects are among them, which is the pipeline that eventually reaches allied health, and it means the tools you are offered in a few years should arrive with Australian validation data rather than borrowed overseas evidence. Read →
🌏 Global | AI Regulation — On 6 August a diagnostic AI platform was cleared alongside a pre-approved plan governing how the product may change without a fresh submission each time. It is a small piece of regulatory plumbing with one useful lesson for you: "cleared" describes a moment in the past, so the question to put to any vendor is what happens the next time the model is updated underneath you. Read →