An Adelaide hospital trial of an AI scribe in emergency and urgent care has produced one of the first hard Australian numbers on how much of the machine’s work clinicians will actually keep: 58 per cent of AI-generated notes were accepted without changes. The rest were edited, and researchers were blunt about why that matters, warning that AI hallucinations mean human oversight in clinical documentation is not optional.
The figure was reported by Health Services Daily on 2 July, drawing on an evaluation of the scribe running inside the Central Adelaide Local Health Network (CALHN). It is a rare result for the sector: a real-world hospital trial that puts a percentage on clinician trust while refusing to gloss over the technology’s failure mode.
What CALHN is running, and where
The scribe sits in emergency and urgent care, capturing the conversation between a consenting patient and their clinician through a small microphone and turning it into a draft clinical note. According to CALHN, the tool started at The Queen Elizabeth Hospital emergency department and, after positive early feedback, expanded to more clinicians and to the nearby Sefton Park Urgent Care Hub.
Roughly 500 patients are expected to take part in the current research phase over about six weeks, with the results then analysed to judge whether the tool genuinely helps emergency and urgent care. Crucially, CALHN describes its system as purpose-built and validated against traditionally written documentation for safety, developed with the Australian Institute for Machine Learning and the Commission on Excellence and Innovation in Health, with funding from Health Translation SA and the Health Services Charitable Gifts Board.
That detail is worth pausing on. The Health Services Daily headline framed the trial around Lyrebird, the Melbourne-founded scribe now widely used across Australian general practice, while CALHN’s own materials describe a bespoke tool trained on SA Health data. Whichever engine is under the bonnet, the finding is the same: a clinician still has to read every note before it becomes part of the record.
Dr Michael Edmonds, Head of Unit at The Queen Elizabeth Hospital emergency department and among the first to use the tool, told The Indian Sun the technology could free up clinicians and make healthcare delivery faster and more efficient. South Australian Health Minister Chris Picton called AI scribes a potential “game-changer” for cutting the administrative load on doctors and nurses so they can spend less time at the computer.
Why the hallucination caveat is the story
A 58 per cent acceptance rate is not a failing grade. It says more than four in ten notes needed a clinician’s hand, which is exactly what a safety-conscious rollout should expect during evaluation. The number people should watch is the other one the researchers volunteered: the risk that a generative model will confidently invent something that was never said.
The failure mode is well documented in Australian primary care. Guidance surfaced in coverage of AI scribes describes a case where a doctor’s note about a patient’s hands, feet and mouth was rendered by the AI as a diagnosis of hand, foot and mouth disease. Hallucinations are rare in day-to-day use, but in a clinical record a rare error can be a serious one, which is why the Royal Australian College of General Practitioners stresses that clinicians must review scribe output before it is committed and remain medicolegally responsible for it.
An emergency department raises the stakes further. Consultations are fast, interruptions are constant, and the pressure to clear a note quickly is exactly the pressure that makes a clinician more likely to wave through a plausible-looking draft. The value of the CALHN trial is that it measured acceptance in that environment rather than in a quiet clinic, and still landed on a message of mandatory human review.
The Australian and Australasian picture
South Australia is not experimenting in isolation. The Indian Sun reported that the state’s AI push is backed by a dedicated healthcare AI trials stream inside a $28 million Digital Investment Fund allocation, with a parallel trial of a commercial scribe underway in the Women’s and Children’s Health Network. Commission on Excellence and Innovation in Health chief executive Michael Brown framed the goal simply: less admin means more care delivered, faster.
Across the Tasman, Health New Zealand’s ED trials offer a preview of the upside at scale. RNZ reported that an AI scribe cut after-hours administrative work by 81 per cent and let doctors see on average one additional patient per shift, with a planned expansion to 14 more emergency departments. Even there, Health NZ’s digital innovation director acknowledged the tool did not get everything right and that doctors had to check it, echoing the Adelaide caveat exactly.
The New Zealand programme also surfaced a concern the Adelaide numbers cannot capture: patient trust. A primary health academic warned that if patients do not trust the consultation, they will hold things back. That is why consent, not just accuracy, sits at the centre of Australian scribe guidance, and why CALHN’s insistence on consenting patients and onshore, validated processing is more than a compliance footnote.
Why it matters for Australia
Health systems in New South Wales, Victoria and beyond are weighing far larger scribe rollouts, and until now they have had vendor demos and overseas studies more than local hospital evidence. The Adelaide trial gives them two things at once: a credible upside for clinician time, and a concrete acceptance rate that sets a realistic expectation rather than a marketing promise.
The 58 per cent figure is the more useful number for a procurement team than any efficiency claim. It reframes the AI scribe not as an autopilot but as a drafting assistant whose output must be checked, and it lets a health network model the review burden honestly before signing a contract. A tool that produces a usable first draft most of the time, and demands editing the rest of the time, is a genuine productivity gain only if the workflow keeps a clinician firmly in the loop.
It also sharpens the governance question. Purpose-built systems trained on local data, validated against existing documentation and processed onshore give a health service more control over safety and privacy than a generic commercial product, but they cost more to build. SA’s decision to co-develop with the Australian Institute for Machine Learning is a bet that the control is worth the price, and the trial results will help other jurisdictions judge whether to buy, build, or wait.
What to watch next
The near-term test is the full analysis of the six-week research phase: not just acceptance rates, but what kinds of errors clinicians caught, whether accuracy holds up in a busy ED, and whether the time savings seen overseas materialise in South Australia. If CALHN can pair a strong acceptance rate with a documented, reliable review step, it will hand the rest of the country a template for scaling AI scribes without quietly outsourcing clinical judgement to a model that sometimes makes things up.
Sources: Health Services Daily, Central Adelaide Local Health Network, The Indian Sun, RNZ, Royal Australian College of General Practitioners.
