It is already reading mammograms, listening in on general practice consultations, drafting clinical notes, and beginning to change the texture of what it means to receive — and deliver — care. The pace has outrun the policy response. And the question of how far AI will go in healthcare now sits alongside a more difficult one: how will the people who use the system have any substantive hand in shaping it?
The radiology reading room: where AI got there first
Medical imaging has been the most visible proving ground for clinical AI. Radiology, pathology, ophthalmology, and dermatology share a common feature: they are built around pattern recognition in visual data, which is precisely what modern deep learning systems are good at. Eric Topol, the cardiologist-scientist whose 2019 book Deep Medicine remains one of the most widely cited statements on AI in clinical practice, has argued that these pattern-recognition specialties would be the first to feel AI’s influence — not through replacement, but through augmentation of human interpretation (Topol, 2019).
The evidence is now catching up with the thesis. A 2025 study in Radiology, reporting on the RSNA Screening Mammography Breast Cancer Detection AI Challenge, found that the top-performing algorithms achieved accuracy comparable to average screening radiologists in Europe and Australia (Chen et al., 2025). Broader reviews confirm that AI can detect subtle markers — microcalcifications, architectural distortions — with increasing sensitivity, and that when combined with human readers, AI-assisted mammography can reduce reading time while lifting cancer detection rates.
But diagnostic performance on test sets is not the same as real-world integration. Australian research published in Cancer in 2025 examined the preferences of women eligible for breast cancer screening. The findings were telling: strong preferences against AI being used alone or as a triage tool, and a population split into three groups — positive about AI if accuracy improves (40%), strongly against it (42%), and concerned about it (18%) (Pearce et al., 2025). Technical readiness is moving faster than social readiness. That gap is where the trust question lives.
AI in the consulting room: ambient scribes and Australian primary care
If radiology is where AI made its first visible impact, general practice is where it is now moving fastest. Ambient AI scribes — systems that listen to a consultation and automatically generate draft clinical notes — have moved from pilot to routine use in a matter of two years. Two Australian companies, Heidi Health and Lyrebird Health, now support a significant share of this market, with Heidi alone reportedly supporting more than one million consultations per week and both companies integrating directly with the major practice management platforms (Pulse+IT, 2025; Medical Republic, 2025).
The productivity effect on clinicians is substantial. Practices reporting real-world use have described saving 5–20 minutes per patient and reducing documentation workload by up to 40% (RACGP, 2025). For clinicians drowning in after-hours paperwork — a well-documented driver of burnout — the effect is meaningful. The clearer value, though, is upstream: if documentation compresses from hours to minutes, attention returns to the patient in front of the clinician. That is the Deep Medicine premise in everyday action: freeing clinicians from keyboards so they can do the work only humans can do.
The technology is not without risks. The Australian Commission on Safety and Quality in Health Care released a dedicated AI Safety Scenario: Ambient Scribe in August 2025, warning clinicians that ambient scribes use generative AI to summarise rather than transcribe verbatim — meaning the risk of hallucinations, omissions, and bias is real and ongoing. Responsibility for the accuracy of the record remains with the clinician, who must review each summary before it enters the healthcare record (ACSQHC, 2025). The TGA has also signalled that tools going beyond transcription will increasingly be treated as clinical technologies, not productivity aids, with regulatory implications to match (ACS Information Age, 2025).
What changes when the administrative layer shrinks
Taken together, the radiology and primary care shifts point to something larger than efficiency. They point to a redistribution of clinical attention.
Topol’s central argument in Deep Medicine is that modern medicine has become “shallow”: short consultations, screen-facing clinicians, a relentless documentation load, and a corresponding erosion of the trust that makes care work. His case is that AI’s most important contribution will not be diagnostic brilliance but the recovery of time — time for clinicians to listen, to think, and to connect. As Topol has put it in interview, machines may eventually out-detect and out-diagnose clinicians, but they will not out-care them (Topol, 2021).
Translating that into an Australian service design frame, at least three shifts follow:
1. Task redistribution, not workforce replacement. The credible near-term trajectory is that AI takes on pattern-matching and documentation tasks while human clinicians do more judgment, communication, and complex care.
2. Geography becomes less binding. AI-assisted triage, remote image interpretation, and asynchronous review have real implications for rural and remote service models — if the infrastructure and workforce policy catch up.
3. The data substrate becomes the constraint. Models are only as good as the populations they were trained on. Systems built predominantly on Northern Hemisphere data will embed bias when deployed in Australia, particularly for Aboriginal and Torres Strait Islander populations, culturally and linguistically diverse communities, and rural cohorts underrepresented in training data.
How far can AI go? The honest answer
Public conversation about AI in healthcare tends to oscillate between two framings: dystopia (machines replacing doctors, patient safety compromised) and uncritical optimism (AI solves the workforce crisis). Neither is a useful guide for system design.
The more grounded answer is that AI’s role will be extensive but bounded. Pattern recognition tasks — imaging interpretation, triage, documentation, risk stratification — will increasingly be AI-augmented. Judgment tasks that depend on contextual reasoning, ethical weighing, and the therapeutic relationship will remain human. The boundary between those two categories is not fixed; it will move with the technology, the regulation, and the evidence base.
Even on the diagnostic side, current evidence suggests limits. Recent comparison studies show that while AI can outperform radiologists on specificity in non-dense breast tissue, radiologists remain more accurate overall in dense breasts — and catch errors the AI misses (ScienceDirect, 2025). The strongest performance, repeatedly, comes from human-AI combination rather than either acting alone.
The trust question
The Pearce et al. (2025) finding — that a substantial share of Australian women are hesitant about AI in breast cancer screening — is not an isolated data point. The 2026 Edelman Trust Barometer shows a broader pattern of global insularity, with trust in institutions concentrated in immediate circles and a growing gap between expectation and performance for leaders (Edelman, 2026). Healthcare is one of the highest-trust sectors in most countries, which is an asset. It is also easily squandered if AI is introduced as a fait accompli rather than as a shared project.
Critically, Australian work has begun to ask the public directly what they think. The 2024 citizens’ jury on AI in Australian healthcare brought together a panel of diverse Australians to deliberate on how AI should be used. Their recommendations — which emphasised transparency, human oversight, and equitable access — align closely with the governance direction being recommended by the Australian Alliance for Artificial Intelligence in Healthcare (AAAiH), led by Professor Enrico Coiera at Macquarie University’s Australian Institute of Health Innovation. The AAAiH’s 2023 National Policy Roadmap for AI in Healthcare calls for a National AI in Healthcare Council, minimum AI safety and quality practice standards tied to accreditation, and explicit attention to the risks of generative AI in clinical documentation (AAAiH, 2023).
Co-design as the trust pathway
The evidence is unambiguous that trust in clinical AI is built — or lost — during the design process, not after deployment. A widely cited 2022 framework in Patterns by Banerjee and colleagues made this case directly: AI algorithms in healthcare should be co-designed with patients, and patients should be involved in decisions about when and how AI research is applied in care (Banerjee et al., 2022). The framework proposes research advisory groups that include people with lived experience, structured walk-throughs of how models are built, and explicit mechanisms to set realistic expectations on both sides.
This dovetails with the broader Australian co-design literature. KA McKercher, in Beyond Sticky Notes: Co-design for Real, sets out four principles that translate almost directly to the AI context: share power, prioritise relationships, use participatory means, and build capability (McKercher, 2020). Applied to AI in healthcare, these principles demand more than a consumer representative on a steering committee.
1. Share power. Consumers and clinicians should have genuine decision rights over whether an AI tool is deployed, how it is configured, and under what conditions it is paused or withdrawn.
2. Prioritise relationships. Building trust takes repeated, transparent engagement — not a single consultation round before procurement.
3. Use participatory means. Community panels, citizens’ juries, and lived-experience research advisory groups give the public a structured voice at each stage of the AI lifecycle.
4. Build capability. Both consumers and clinicians need enough literacy about AI to participate meaningfully — otherwise “co-design” becomes consent to something people don’t really understand.
Work published in the Croakey health media has made a similar case: the choices being made now about digital health infrastructure — interoperability, privacy, AI decision support — will determine whether AI in Australian healthcare is something done to consumers or designed with them (Croakey, 2026).
Governance as infrastructure, not paperwork
Co-design only goes as far as governance allows. The current regulatory environment is catching up. The Commission’s 2025 AI Clinical Use Guide sits alongside the TGA’s stepped-up oversight of AI scribes as medical devices, the Australian Government’s proposed mandatory guardrails for AI in high-risk settings, and the sector-specific work of the AAAiH. These layers are beginning to add up to something that resembles a national architecture. What remains missing is the coordinating mechanism that the AAAiH has been calling for — an authoritative national body that can set standards, respond to incidents, and anchor public confidence across jurisdictions (Coiera, 2023).
Implications for practitioners
For consultants, commissioners, and service designers working in Australian health and social care, the practical implications of the AI turn are already reshaping the work.
Business cases increasingly have an AI dimension. Workforce, productivity, and service redesign proposals now need to address where AI sits within the model — including what assumptions are being made about task redistribution, data infrastructure, and clinician oversight.
Evaluation frameworks need to catch up. Evaluating an AI-enabled service requires measuring not only accuracy and efficiency, but equity (who is and isn’t served well), trust (do users understand and accept how AI is being used), and workforce impact (is AI shifting burden, or adding to it?).
Co-design is no longer optional. For any AI-enabled care model to earn public trust, consumers and clinicians need to be in the room before deployment, not asked to validate it afterwards. This is where the evaluation, commissioning, and co-design disciplines converge.
At Ethicol, this work sits at the intersection of evaluation methodology, co-design practice, and policy analysis. The AI transformation of healthcare is already underway. The question is not whether to engage with it, but how to engage in a way that keeps people — patients, clinicians, and communities — genuinely at the centre of the system being built.
The centre holds, or it doesn’t
AI in healthcare will go as far as Australians let it — and trust it — to go. The radiology, scribe, and triage applications now in use are the beginning of a much broader shift in how care is produced. The technical trajectory is clear. The social one is not, and will not resolve itself.
The organisations that will navigate this period well won’t be the ones that adopt AI fastest. They will be the ones that adopt it with the deepest and earliest engagement from the people who use their services. If the AI turn in healthcare leaves consumers out of the design process, trust will erode faster than efficiency can compensate for. If consumers are genuinely at the centre of the design, AI has a real chance to do what Topol argued it could: make healthcare more human, not less.
The centre holds, or it doesn’t. That is a design choice.
References
Australian Alliance for Artificial Intelligence in Healthcare [AAAiH] (2023). A National Policy Roadmap for Artificial Intelligence in Healthcare. Sydney: AAAiH / Macquarie University. Available at: aihealthalliance.org
Australian Commission on Safety and Quality in Health Care [ACSQHC] (2025). AI Safety Scenario: Ambient Scribe, Version 1.0, August 2025. Sydney: ACSQHC.
Banerjee, S., Alsop, P., Jones, L., et al. (2022). Patient and public involvement to build trust in artificial intelligence: A framework, tools, and case studies. Patterns, 3(6), 100506.
Chen, Y., Partridge, G. J. W., Vazirabad, M., et al. (2025). Performance of algorithms submitted in the 2023 RSNA screening mammography breast cancer detection AI challenge. Radiology, 316(2), e241447.
Coiera, E. (2023). Remarks on the launch of the National Policy Roadmap for AI in Healthcare. AI.Care conference, Melbourne, 22 November 2023.
Croakey Health Media (2026). From users to co-designers of Australia’s digital health future. Croakey, February 2026.
Edelman (2026). 2026 Edelman Trust Barometer Global Report. Available at: https://www.edelman.com/trust/2026/trust-barometer
McKercher, K. A. (2020). Beyond Sticky Notes: Co-design for Real — Mindsets, methods and movements. Sydney: Beyond Sticky Notes.
Pearce, A., Tuffaha, H., Scuffham, P., et al. (2025). Implementing artificial intelligence in breast cancer screening: Women’s preferences. Cancer. Published online April 2025.
Royal Australian College of General Practitioners [RACGP] (2025). Artificial intelligence (AI) scribes in general practice. RACGP running-a-practice technology resource.
Topol, E. (2019). Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. New York: Basic Books.
Topol, E. (2021). Remarks, NIH Director’s Wednesday Afternoon Lecture Series, September 2021.
.avif)
.avif)
.avif)
