Queensland Health widened a generative-artificial-intelligence clinical documentation trial in emergency departments on Thursday, deploying ambient “scribe” software that drafts encounter notes from clinician-patient conversations after the Therapeutic Goods Administration (TGA) updated expectations for software classified as a medical device. Hamish Clarke’s AI desk tracks the gap between vendor demos and what a Commonwealth regulator will actually allow: here, AI may listen and propose text, but a licensed clinician must review and sign before anything lands in the legal record or flows to My Health Record.
What the trial does in ED bays
Microphones on secured tablets capture de-identified audio in treatment spaces where patients consent via Queensland’s standard digital-care notices. A foundation model tuned on Australian medical terminology produces draft histories, examination findings, and disposition plans formatted for the state’s electronic medical record. The Guardian Australia reported that Royal Brisbane and Women’s Hospital and Cairns Hospital joined earlier sites in Logan and Gold Coast, doubling the number of overnight shifts covered during the September respiratory season bump.
Queensland Health said the scribe does not recommend diagnoses or order medications autonomously—it accelerates typing for staff facing queue backlogs. That boundary matters for TGA classification: decision-support versus autonomous practice.
TGA lens on software as a medical device
The TGA’s refreshed guidance stresses risk classification for models that influence clinical management, even when sold as documentation aids. Vendors must show cybersecurity, bias testing, and update controls when weights change. Queensland’s procurement required conformity assessments for the scribe platform before microphones switched on in new bays—Clarke notes that many startups pitch “administrative AI” to avoid device rules; regulators are closing that loophole when outputs touch patient charts.
Federal health minister spokespeople have said states retain delivery accountability; the TGA sets the guardrails. If a scribe hallucinates an allergy or medication dose in draft form, the harm is contained only if clinicians catch errors—Queensland reports mandatory diff highlighting for mismatches against structured vitals feeds.
Privacy, consent, and workforce politics
Australian Digital Health Agency policies on My Health Record still require patient visibility into shared documents. AI drafts are not uploaded until signed, reducing the risk of premature disclosure. Nurses’ unions asked for workload impact studies: if scribes save physicians ten minutes per encounter but add nurse verification steps, net benefit is unclear. Queensland promised quarterly metrics on time-to-discharge documentation and clinician satisfaction, published without patient identifiers.
Patients may opt out of audio capture; fallback is manual typing. In busy trauma bays, ambient listening is paused for sensitive conversations by policy, not only by technology limits. Interpreter services remain human-led; the scribe does not auto-translate unless a separate certified module is approved.
Comparison with other states
New South Wales and Victoria run smaller documentation pilots with different vendors; none yet match Queensland’s ED scale. Western Australia is watching TGA rulings before funding statewide contracts. Clarke’s desk expects a patchwork unless the Australian Health Ministers’ Advisory Council harmonises procurement—unlikely before federal election timing firms up.
Private hospital groups experiment with their own scribes under insurer scrutiny; public trials carry higher transparency obligations and union engagement.
Evaluation criteria through year-end
Queensland Health will compare error rates in signed notes against historical baselines, audit a random sample for fabricated statements, and track medico-legal incidents. Vendor models may not train on Queensland patient audio without separate ethics approval—a line Clarke repeats because US vendors often default to broader data rights. Independent clinicians on the trial board can halt deployments site by site without waiting for a statewide decision.
If metrics hold, the state may tender a single scribe platform for 2027 ED budgets; if not, microphones go dark regardless of hype. Thursday’s expansion is policy-meets-product: TGA clarity gave cover to scale, but the clinical sign-off rule means AI is still an intern, not a registrar. Rural EDs on the pilot list will test satellite-backhauled sites separately so documentation delays are not blamed on bandwidth alone.
