AI scribes: Australian patients weigh communication over accuracy

When the Medical Journal of Australia asked 275 patients what they thought of doctors using AI scribes to generate their consultation summaries, the answers did not split along the lines the health-tech industry has been bracing for. Patients were not outraged. They were not demanding the software be removed from clinics. But nor were they reassured by assurances about transcription accuracy, data encryption or clinician-in-the-loop design. They wanted something simpler, and harder to deliver: evidence that the AI had understood the consultation the way a human would.

The survey, authored by Monash University researchers Wei Zhou, Rashina Hoda and Joycelyn Ling and published in MJA InSight+ this month, asked participants to evaluate AI-generated after-visit summaries against the same summaries edited by a clinician. Patients consistently preferred the edited versions — and not because they were factually more correct. The unedited AI summaries were accurate. The problem was tone.

This summary feels more patient-centred as it feels like less of a statistic dump and more of an empathetic summary of the consultation with less egregious data.
— Survey respondent, quoted in MJA InSight+

Another respondent put it plainly: “If AI summaries and tools were required to be checked over by healthcare professionals before patients had access to them, I would be more likely to use them.”

The finding lands at a moment when AI scribes have gone from pilot to procurement-line-item in Australian general practice. A Healthed survey of 1,535 GPs published this month found 18.7 per cent of respondents personally used an AI scribe for consulting, with 37.3 per cent of those users deploying the tool in 80 to 100 per cent of consultations. That is a narrower number than the 40 per cent figure from a November 2025 RACGP poll cited by The Guardian, but both datasets make the same point: the technology is spreading faster than the governance frameworks meant to contain it. Only one in four Australian general practices has an AI governance policy, according to research cited by the MJA authors.

Melbourne-based Heidi Health has ridden the wave to a $660 million valuation and annualised revenue of $US50 million ($76 million), up from $US1 million two years ago, the Australian Financial Review reported in April. Copenhagen-based Corti this month launched a specialised speech-to-text model — Symphony — that it claims beats OpenAI’s general-purpose models on medical terminology accuracy. The market is growing and the technology is improving. But the patient survey suggests the evaluation framework the industry uses to measure that improvement — word-error rates, diagnostic coding accuracy, note completeness — misses the thing patients are actually judging.

The consent story is no tidier. In reporting for AusDoc, journalists found clinics using waiting-room posters, offhand verbal agreements at the start of a consultation, and in some cases, no explicit consent mechanism at all. Consumer Health Forum chief executive Elizabeth Deveny told The Guardian the power imbalance between a patient and the clinician they depend on makes meaningful opt-out difficult: people do not want to be seen as difficult. The MJA study itself noted that Australian patients are rarely asked whether they are comfortable with AI scribe use before a consultation begins.

If AI summaries and tools were required to be checked over by healthcare professionals before patients had access to them, I would be more likely to use them.
— Survey respondent, quoted in MJA InSight+

This is a regulatory gap with teeth. AI scribes operate in Australia without specific Therapeutic Goods Administration oversight — they are exempted as administrative tools rather than clinical decision-support software. That classification is defensible so long as the tool remains a passive transcription engine. But the Monash study makes clear that patients experience the output as clinically consequential: the summary shapes how they understand their own health. When tone, warmth and empathy become the metrics patients use to evaluate quality, the line between administrative and clinical blurs.

Then there is the cognitive question. Caitlin Curtis, an AI ethics researcher at the University of Queensland, has warned that outsourcing clinical note-taking removes a process that helps doctors reflect on what they heard and retain the nuance of a consultation. The burnout numbers complicate the picture: a JAMA Network Open study found that doctors meeting burnout criteria dropped from 51 per cent to 29 per cent within 42 days of introducing AI scribes. A less exhausted GP is almost certainly a better GP. But Curtis’s question — does the machine-written record degrade clinical memory over time? — has no published answer yet.

The MJA survey does not settle any of these tensions. It does something more useful: it tells the health-tech sector what patients would need to see before they trust the software sitting between them and their doctor. Respondents wanted summaries that sounded like a person wrote them. They wanted proof a clinician had reviewed the output, not an assurance buried in a privacy policy. And they wanted to be told, before the consult began, that an AI would be listening.

AI is cool but medical practice cannot 100 per cent depend on AI — AI is the assistance and functional extension of human medical experts.
— Survey respondent, quoted in MJA InSight+

That posture — AI-as-assistant, demonstrably supervised, evaluated on empathy — is not what the procurement checklists currently measure. The Monash team’s call for a patient-centred evaluation framework is not a research-niche ask. It is a read on where the rollout’s legitimacy will crack if nobody acts on it.