AI is not replacing doctors, and on current evidence it will not in the foreseeable future. It is replacing specific tasks within medical work: reading, sorting, drafting and flagging. Those tasks are spread unevenly across specialties, so image-heavy, high-volume, repeatable work changes first, while the job itself, owning a patient’s decision, stays human.
In 2016 the computer scientist Geoffrey Hinton, later a Nobel laureate, told an audience that people should stop training radiologists, because deep learning would soon read scans better than they could. It became the most quoted prediction in medical AI.
A decade on, the United Kingdom does not have too many radiologists. It has far too few. The Royal College of Radiologists’ 2024 workforce census found a 29% shortfall of clinical radiologists, forecast to reach 39% by 2029 without action, as demand for CT and MRI grew faster than the workforce. Meanwhile radiology absorbed more AI than any other field. By the end of 2025 the US Food and Drug Administration had authorized 1,451 AI-enabled medical devices, and 76% of them were for radiology.
So the specialty with the most AI also has one of the worst staffing gaps. That is not a paradox once you stop asking the wrong question.
“Will AI replace doctors?” treats a doctor as a single unit of labor that a machine either can or cannot do. A doctor is actually a bundle of dozens of tasks, and AI is eating that bundle from the edges, one task at a time, at very different speeds depending on the specialty. The useful question is which tasks go, which get harder, and who owns the decision when the easy ones disappear.
Which medical tasks is AI actually taking over?
Look at what the regulated tools do. The FDA’s list of AI-enabled devices is dominated by software that detects, measures, triages or prioritizes: a suspected bleed pushed up the reading queue, a lung nodule measured, a mammogram flagged as low risk. These are perception tasks with a clear input, a clear output and plenty of labeled training data.
Outside imaging, the fastest adoption is in paperwork, not diagnosis. The American Medical Association’s physician survey, fielded in January 2026, found that more than 80% of responding physicians used AI in a professional context, double the share in 2023, with most use centered on documentation and summarization. Ambient scribes, discharge summaries, prior authorization letters and literature summaries are where the hours are being clawed back.
Put those two lists together and a pattern appears. AI takes tasks that are high volume, repeatable, well defined and checkable after the fact. It struggles with tasks that are rare, ambiguous, negotiated with a patient, or where nobody agrees on the right answer.
That means exposure varies enormously. A screening radiologist reading thousands of mostly normal mammograms has a large share of work that fits the first description. A palliative care physician spending an hour on a goals-of-care conversation has almost none. A dermatologist doing skin checks sits somewhere in between: the lesion photograph is automatable, the biopsy decision and the conversation are not.
Why hasn’t AI replaced radiologists yet?
Because the best evidence shows AI changing the shape of radiology work rather than removing the radiologist.
The strongest trial is MASAI in Sweden, which randomized more than 100,000 women to AI-supported mammography screening or standard double reading. Interim results in The Lancet Oncology in 2023 showed a 44% reduction in screen-reading workload. The full results, published in The Lancet in January 2026, found that the AI-supported arm detected more cancers at screening and had 12% fewer interval cancers, with a similar false-positive rate.
Read the design carefully. The AI did not read the mammograms alone. It triaged them: low-risk exams went to a single radiologist, high-risk exams to two, and the highest-risk cases were flagged. The trial team was explicit that AI-supported screening still requires at least one human radiologist to read every screen.
That is task replacement in its clearest form. The second reader on a normal mammogram, a task that consumed a large share of screening radiologist time, largely went away. The radiologist did not. And the 44% of workload freed up in a country short of breast radiologists does not become unemployment. It becomes capacity for the backlog, for diagnostic workups, for the cases AI flagged.
This is also why the evidence for AI diagnostics looks strongest where the clinical question is narrowest, a point made in our earlier piece on where AI diagnostic evidence holds up. A narrow question can be automated. A whole specialty cannot.
Which medical specialties will AI change first?
A simple test predicts exposure: what fraction of the specialty’s working hours go to tasks that are high volume, standardized and produce an output a second person could verify?
By that test, the early movers are screening radiology, pathology slide review, ophthalmology screening for diabetic retinopathy, dermatology photo triage and ECG interpretation. In each, a large block of time goes to looking at a standardized image or signal and deciding whether it is normal.
The next tier is documentation-heavy specialties, which is nearly all of them. Primary care physicians, psychiatrists and hospitalists spend a large share of their day writing. Scribes and summarizers do not replace any clinical judgment, but they change what a working day consists of.
The least exposed are specialties built around procedures, uncertainty and relationships: surgery, emergency resuscitation, obstetrics, palliative care, pediatrics at the bedside. AI may assist with planning or monitoring, but the task that defines the job requires hands, presence or negotiated judgment.
The implication for workforce planners is uncomfortable. The specialties most exposed to AI are not overstaffed. Radiology and pathology are among the tightest labor markets in medicine. That makes AI more likely to fill gaps than to create redundancies in the near term, and it means any savings will show up as throughput, not headcount.
A worked example: one screening radiologist’s week
Consider a hypothetical breast radiologist in a public screening program, with round numbers chosen to show the shape of the change.
Before AI, she spends about 20 hours a week double-reading screening mammograms, 10 hours on diagnostic workups and biopsies, 5 hours in multidisciplinary meetings and 5 hours on administration and teaching.
Now the program adopts AI triage modeled on the MASAI approach. Low-risk exams need one reader instead of two. Suppose her screening reading time falls by roughly 40%, in line with the trial’s workload reduction, from 20 hours to 12.
What happens to the 8 hours? In a service with a waiting list for diagnostic workups, they go straight into biopsies and recall clinics. In a service with an AI-flagged queue of high-risk cases, some go into harder reads that now arrive concentrated rather than diluted among normals. Her week is more complex per hour, not emptier.
Two things have changed, though. First, the skill mix of her work shifts toward the difficult end, which is more tiring and needs protected time for reading normal cases so that her baseline does not drift. Second, if the program’s managers treat the 8 hours as a saving and cut a post, the remaining radiologists absorb the concentrated hard cases with no slack. The technology is the same in both scenarios. The workforce outcome depends entirely on what management does with the freed hours.
What is the strongest case that AI will replace doctors?
The argument deserves to be stated at full strength.
Tasks compound. If AI takes the reading, then the drafting, then the triage, then the routine follow-up, the residual human job may shrink to a thin layer of sign-off. Sign-off does not need a physician. It could be done by a nurse practitioner, a physician associate or a technologist supervised remotely. Health systems facing cost pressure will be tempted to redraw roles around what is left, and many already use task-shifting where the evidence supports it, as we argued in our piece on task-shifting.
This is a real risk, and it is more plausible than wholesale replacement. The likely path is not “AI replaces doctors” but “AI plus a cheaper professional replaces some doctor hours.” In health systems short of physicians, that may be exactly what patients need. In health systems looking for savings, it may erode supervision in ways that only show up in rare, serious errors.
But there are limits the strong case underestimates. Someone must be accountable for a clinical decision, and licensing, liability and reimbursement all attach to a named professional. Patients still want to hear serious news from a person. And the rare, ambiguous, multi-problem patient, who accounts for a disproportionate share of risk and cost, is exactly where AI performs least reliably. The bundle shrinks, but the core does not disappear.
What this means for doctors, health systems and trainees
For practicing physicians, the practical move is to know which of your tasks are automatable and to get involved in how that automation is deployed. The AMA’s 2026 survey found 85% of physicians want to be consulted on or responsible for AI adoption in their practice. That influence is available now, while workflows are still being designed, and much harder to win later.
For health systems, the decision that matters is not whether to buy AI. It is what to do with the hours it frees. Treat them as capacity for backlogs, harder cases and supervision, and AI extends a scarce workforce. Treat them as a headcount saving, and you concentrate difficulty on fewer people and lose the slack that catches errors.
For medical schools and residency programs, the priority is to protect foundational skills while AI takes routine volume. A trainee who never reads a normal study unaided may never develop the pattern sense to know when the machine is wrong, which is why we have argued that programs should teach residents to work without AI first.
For workforce planners and ministries, the lesson from radiology is that AI does not reliably reduce the number of specialists needed. It raises what each specialist can do. Workforce plans that bank on AI to close gaps before the evidence shows it will are betting patients’ access on a forecast that has already been wrong once.
Frequently asked questions
Will AI replace radiologists?
Not on current evidence. Radiology has the most AI tools of any specialty, yet the UK’s Royal College of Radiologists reports a 29% shortfall of clinical radiologists. The MASAI trial showed AI can cut screen-reading workload by 44% while improving detection, but every screen still had at least one human reader. AI is changing radiology work, not removing the radiologist.
Which doctors are most likely to be affected by AI?
Specialists whose time is dominated by high-volume, standardized interpretation are most exposed: screening radiology, pathology, retinal screening, dermatology triage and ECG reading. Documentation-heavy specialties such as primary care will see large changes in how the day is spent. Procedure-based and relationship-based work, including surgery, emergency care and palliative care, is least exposed.
Are doctors actually using AI already?
Yes, widely. The American Medical Association’s 2026 physician survey, fielded in January and February 2026, found that more than 80% of responding physicians use AI professionally, double the 2023 share. Most use is for documentation and summarization rather than diagnosis, and the number of use cases per physician has roughly doubled since 2023.
Should students still become doctors if AI is improving?
The evidence suggests yes. Physician shortages are widespread, demand is rising with ageing populations, and AI so far extends what clinicians can do rather than replacing them. What changes is the content of the job: more oversight of machine outputs, more complex cases, and a premium on judgment, communication and the ability to recognize when the tool is wrong.
Can AI make a diagnosis without a doctor?
Some regulated tools make narrow determinations autonomously, and consumer chatbots answer health questions for millions of people. But in most health systems, a licensed professional remains accountable for diagnosis and treatment decisions. Anyone concerned about a symptom should consult a clinician rather than rely on an AI tool alone.
The machines are not coming for the doctor. They are coming for the parts of the doctor’s day that looked most like a production line, and what happens next is a management decision, not a technical one.
Which task in your own clinical week would you hand to AI tomorrow, and which one would you never give up?
