--- title: "Replacing doctors is the wrong question" author: Dr Arman Ouveysi date: 2026-07-07 room: Lab tags: medical-ai, automation, future-of-work, ai-engineering canonical: https://drouveysi.com.au/essays/replacing-doctors-is-the-wrong-question/ summary: "Will AI replace doctors is the question I get asked most and it cannot be answered as asked, because "doctor" is not a task. Unbundle the job and the real questions appear, along with some uncomfortable honesty about my own." --- # Replacing doctors is the wrong question Once people learn how my week is split, one half in a consulting room and the other building medical AI, there is a question that arrives with the reliability of a tide: *will AI replace doctors?* It comes in two tones. From patients and friends, it is asked with worry. From certain corners of the tech world, it is asked with relish. Both versions expect a yes or a no, and I have stopped giving either, because the question cannot be answered in the form it is asked. It is like asking whether machines will replace transport. Which part? "Doctor" is not a task. **It is a bundle, and the bundle unties.** Inside it sits information retrieval: knowing the guideline, the interaction, the dose adjustment for the failing kidney. Pattern recognition: the rash, the tracing, the shape of a story that does not add up. Procedures: hands doing things to bodies. Judgement under uncertainty: deciding at 6pm with two-thirds of the picture, because the last third takes a week to arrive and the patient is here now. Accountability: a named human whose registration, insurance and sleep are attached to the decision. And presence: someone in the room while the bad news lands. These components have wildly different exposure to automation, and any sentence that treats them as one thing will be wrong about most of them. Medicine has run this experiment before, quietly, and the result is instructive. Machines have been printing an automated interpretation across the top of ECGs for roughly fifty years. The algorithm is genuinely useful, routinely wrong in characteristic ways, and every doctor is taught to read the tracing first and the machine's opinion second. Nobody remembers a crisis about it. What happened instead is that the job changed shape: the machine absorbed the legible part of the task, the humans concentrated on the illegible remainder, and cardiology as a profession got larger, not smaller. That is the standard trajectory of clinical automation to date. **Not replacement. Redistribution**, with the boundary between machine work and human work migrating one legible task at a time. So the honest version of the question is: which parts of the bundle are migrating now? And the honest answer is: more than my profession is comfortable saying out loud. Documentation is already going, and good riddance; the profession's most hated task turned out to be the most automatable. First-draft synthesis is going: differential lists, guideline retrieval, the summarisation of a fifty-page history into the two paragraphs that carry the case. Out-of-hours triage of the worried-well is going, and this one deserves a moment of candour, because the sentimental objection is that a frightened person at 3am needs a human. The empirical observation is that the frightened person at 3am currently has no human. They have a search engine and four hours of catastrophising, and a system that can take a careful history at that hour, sort the reassurable from the concerning, and say "this needs a doctor today" is not competing with a doctor. **It is competing with nothing, and winning.** And at the far edge, the boundary the field is now walking toward: narrow, rule-shaped clinical decisions, of the kind a routine prescription renewal represents, being worked by machines inside tightly evaluated scopes under a regulatory sandbox, with a clinician authorising every decision while the evidence accumulates. Small, fenced, heavily supervised. Not yet crossed, but chartered. The precedent that matters already exists: a state has built a legal pathway by which a bounded clinical decision could one day be made accountable without a human making it in the moment, and pathways of that kind do not walk backwards. Now the parts that are not migrating, and the reasons, because the reasons are worth more than the list. Accountability does not transfer, and this is structural rather than sentimental. A medical licence is a strange and valuable object: it is a standing promise, backed by a person, enforceable by courts, insurers and registration boards, that someone can be held responsible when a decision goes wrong. You cannot deregister a model. You cannot cross-examine a checkpoint. Every serious deployment of clinical AI therefore ends the same way when you follow the accountability upstream: at humans and institutions who own the system's decisions, define its scope, and answer for its failures. That is not "no doctor in the loop". That is the doctor moving up a level, from making each decision to owning the machinery of decisions, which is a profound change of job description and nothing like disappearance. ![The doctor role as eight parts, each tagged by how AI changes it](https://drouveysi.com.au/assets/articles/lab/replacing-doctors-is-the-wrong-question/fig-01.svg) Embodiment does not transfer on any timeline worth planning around. Examination, procedures, the physical craft of medicine: robotics will chip at the edges for decades before it touches the core. And presence? Here I will disappoint my own profession slightly, because the standard defence, that patients will always need the human touch, is weaker than we want it to be. Plenty of people demonstrably prefer disclosing to a machine; shame is quieter when nobody human is watching, and the 3am confessions these systems receive make that clear. What genuinely does not transfer is not warmth. It is the specific weight of a person taking responsibility in front of you: someone who can be wrong with you, and answer for it. Machines can simulate the bedside manner. They cannot yet stand in the blast radius. The risks worth losing sleep over are not the replacement fantasy. They are two quieter failure modes. The first is deskilling. Aviation learned this decades ago: automate the routine flying and pilots' manual skills atrophy exactly when the automation fails and the skills are needed most, which is why the industry deliberately makes its pilots hand-fly. Medicine is about to run the same experiment on clinical reasoning at scale, with a generation trained to verify machine output, and verification is a skill that decays into rubber-stamping unless it is deliberately exercised. We will need the medical equivalent of hand-flying hours, and nobody has designed them yet. The second is misallocation: the same automation can be spent two ways, buying more attention per patient or more patients per doctor, and the funding models medicine runs on will push hard toward the second. Neither of these problems is technical. Both will determine whether the technology makes medicine better or merely faster. And then there is the reframe that I think outweighs everything above it, which is that the question contains a hidden assumption: that there is a doctor to replace. For billions of people there is not. Roughly half the planet lacks reliable access to essential care; **the binding constraint on medicine was never intelligence, it is attention**, and clinical attention is one of the scarcest commodities on Earth. Automation converts that scarcity into a floor. The ceiling, a good doctor amplified by good machines, rises for the people who already had doctors. The floor rises from nothing to something for the people who never did, and on any honest moral accounting the floor is the bigger story, even though the ceiling gets all the press. Honesty requires one more paragraph, and it is the one my colleagues will like least. Some clinical jobs, as currently shaped, will not survive this, and I include a slice of my own. A meaningful fraction of what I do in a clinic week is rule-shaped: stable patients, established diagnoses, monitoring and renewal, the same careful decision executed many times. That is precisely the work that migrates first, and I know it in unusual detail, because building the systems that absorb such work is my other job. Pretending otherwise would be comfort dressed as analysis. The parts of my clinical work that are hardest to automate, the undifferentiated mess, the multi-system puzzles, the judgement calls with incomplete pictures, are also the parts I would keep if I could keep only some. There is a version of this transition that concentrates doctors on exactly that work. There are also versions that just fire people, and the difference between those futures is policy and design, not capability. So, replace the question. Not *will AI replace doctors*, which unravels on contact, but three questions with actual answers under construction right now: which decisions get automated, and who draws those boundaries; who is accountable when the machine decides; and what do we deliberately preserve in the humans while the machines absorb the legible. Those are the questions I work on all week, in both buildings. The tide question I have learned to answer with a question: which part? Watching someone realise the bundle unties is usually the most useful thing I do at that table.