■ MODERATE RISK ■ Healthcare
No, though the diagnostic monopoly is over. Medicine is being restructured around AI decision support, and the physicians who thrive will be the ones who supervise it well rather than compete with it.
“AI diagnoses better. But patients want a human to deliver the news and hold their hand.”
Our AI replacement risk score — how we score jobs
A physician's day is mostly pattern recognition wrapped in conversation and paperwork. Take a history, examine, order and interpret tests, form a differential, prescribe or refer, document everything, chase results, coordinate with specialists, and repeat every ten to fifteen minutes in primary care. Hospital medicine adds ward rounds, admissions, escalation decisions, and the constant triage of who is sick enough to worry about now.
AI is already strong at several of these. Image interpretation in radiology, dermatology, pathology, and ophthalmology performs at or above specialist level in narrow benchmark tasks. Symptom-checking and differential generation are genuinely useful, particularly for rare presentations a generalist sees once a decade. Ambient documentation tools now write the clinical note from the consultation, which addresses one of the profession's largest time sinks and burnout drivers. Risk scoring, drug interaction checking, and guideline retrieval are all better done by machine. Real deployment is under way, and it changes the work meaningfully.
It doesn't remove the physician, for reasons that are practical rather than sentimental. Patients present with ambiguity: vague symptoms, incomplete stories, things they're embarrassed to say until minute nine of a consultation. Eliciting that history is a skill. Physical examination requires hands. Prescribing authority, admission decisions, and clinical accountability are regulated — a licensed human bears legal responsibility for the outcome, and no regulator is preparing to transfer that to a model. Complex, multimorbid patients require weighing treatments against each other and against what the person actually wants, which is a values conversation. Breaking bad news, negotiating end-of-life care, and persuading someone to change their behaviour are irreducibly human. Our 30 describes augmentation with teeth: fewer physicians may cover more patients, and specialties heavy in image interpretation will feel it most.
Automatability: our editorial assessment of current and near-term AI capability
Augmentation is happening now — ambient documentation and imaging support are in real clinical use, not pilots. By 2030 expect AI decision support to be standard, with diagnostic-heavy specialties absorbing the biggest change to their workflow. Physician numbers are constrained by shortages in most countries, so the near-term effect is more capacity per doctor rather than fewer doctors. Meaningful restructuring of specialty demand is a 2040 question.
It replaces parts of what doctors do, particularly interpretation of images and structured data where models already match specialist performance in narrow tasks. It doesn't replace the licensed clinician who examines the patient, holds legal responsibility for prescribing and admission decisions, and navigates what a patient actually wants when treatments have tradeoffs. Expect supervision of AI rather than substitution.
Those weighted toward interpreting images and data — radiology, pathology, dermatology, and ophthalmology screening — see the largest workflow change. Historically, though, better diagnostic tools have increased demand for the specialists who act on findings rather than reducing it. The least affected are procedural, examination-heavy, and complex-care specialties where hands and judgment dominate.
Yes. Ageing populations, workforce shortages in most health systems, and long training pipelines make physician demand structurally strong. What changes is the skill mix: less value in memorized recall and routine interpretation, more in supervising AI output, procedural competence, and communication. Our risk score of 30 reflects transformation of the work rather than a threat to employment.
Increasingly for triage and screening, considerably less for anything serious. People want a person to explain what a scan means, to be accountable if something is missed, and to sit with them through bad news. Trust is part of the treatment — adherence, disclosure, and honesty in a consultation all depend on a relationship that a clinician builds and software doesn't.