Will AI Replace Physicians?

What the employment figures say

The US Bureau of Labor Statistics counted 112,010 people in its closest category, 29-1215 Family Medicine Physicians, in May 2023, at a mean annual wage of $240,790 and a median of $224,640, and does not attribute those levels to AI ^{(US Bureau of Labor Statistics 2023)}. The official series agrees with the exposure reading rather than contradicting it: employment in that category rose from 98,590 in May 2020 to 102,930 in May 2021 and 112,010 in May 2023, with mean annual wages moving from $214,370 to $235,930 to $240,790, and neither table credits AI for the change ^{(US Bureau of Labor Statistics 2020)}. There is no BLS projection for 29-1215 specifically; the only official forecast covers the broader Physicians and Surgeons category, 839,000 jobs in 2024 with 3% growth projected to 2034 and about 23,600 openings a year, and that page models neither ambient scribes nor diagnostic systems ^{(US Bureau of Labor Statistics 2024)}. Read the category carefully: it is US-only, family medicine is the nearest analogue to a GP rather than a match, and the broader projection folds in many specialties well beyond general practice.

What the research says about physicians

Now

Where AI has arrived in general practice, it has arrived at the keyboard. In a Nuffield Trust and Royal College of General Practitioners survey of 2,108 UK GPs — the RCGP being a professional association with policy and member interests, though the report sells no product — 28% reported using AI, and among those users 57% used it for documentation or note taking, 44% for administrative tasks and 28% for clinical decision support ^{(Nuffield Trust and Royal College of General Practitioners 2025)}. Kaiser Permanente, reporting on its own deployment and with an organisational interest in a positive account, ran a generative-AI ambient scribe across 7,260 physicians and 2,576,627 encounters and counted nearly 16,000 documentation hours saved, with no physician headcount change reported ^{(Kaiser Permanente Division of Research 2025)}. The sharpest counter-evidence is clinical rather than clerical: that 2025 comparison on 200 commonly asked clinical questions put ChatGPT-4o ahead of family physicians on accuracy, appropriateness and empathy at p below .01, while testing none of the physical examination, longitudinal care, statutory sign-off or liability the job actually runs on ^{(International Journal of Medical Informatics 2025)}. Deployment is also gated institutionally, with NHS England publishing guidance on AI-enabled ambient scribing products in April 2025 that makes adoption conditional on authorisation rather than capability alone ^{(NHS England 2025)}.

The next few years

The applicability research points the same way as the survey data: diagnosis and documentation are strongly augmented, while physical examination, licensure and treatment accountability resist ^{(Microsoft Research 2025)}. Adoption numbers are scaling fast — UPMC's rollout of the Abridge ambient documentation platform targets more than 12,000 clinicians by 2026, announced by the vendor that sells the automation and reporting no headcount reduction, layoffs or lowered hiring target ^{(Abridge 2025)}. Pulling hard in the other direction is demand: the AAMC projects a US shortfall of up to 86,000 physicians by 2036 from population growth, ageing, work-hour assumptions and attrition, without modelling AI, and the AAMC is a medical-school and workforce-policy body with an interest in workforce expansion ^{(Association of American Medical Colleges 2024)}. The AAFP, the trade body for family physicians, argues on the same side, advocating for as many as 48,000 additional primary-care physicians by 2034 and citing a possible shortage of 58,000 primary-care clinicians by 2040 ^{(American Academy of Family Physicians n.d.)}.

Longer term

Eurostat counted about 1.98 million practising physicians in the EU in 2023, including about 500,200 generalist medical practitioners, with physician density per 100,000 inhabitants rising in every EU country between 2018 and 2023 ^{(Eurostat 2023)}. Ageing in the profession points to scarcity rather than surplus: at least 39% of doctors were aged 55 or older in ten EU countries, reaching 54.5% in Bulgaria, which Eurostat says can contribute to future shortages ^{(Eurostat 2023)}. Underneath all of it sits a constraint no model release changes, because GMC guidance states that doctors remain responsible for decisions made using AI, so professional accountability stays with the physician even when software supplies the recommendation or the draft ^{(General Medical Council n.d.)}.','

What is exposed, and what protects it

● Where the exposure is

The exposed surface is the paperwork and the well-posed clinical question. Documentation is the single clearest case: an ambient scribe across 7,260 physicians and 2,576,627 encounters produced nearly 16,000 saved documentation hours and less after-hours record work, on the deploying organisation's own account ^{(Kaiser Permanente Division of Research 2025)}. Among GPs already using AI, 57% use it for documentation or note taking and 44% for administrative tasks ^{(Nuffield Trust and Royal College of General Practitioners 2025)}. On text-based clinical reasoning the models are not merely adequate — one 2025 comparison across 200 common clinical questions rated ChatGPT-4o above family physicians on accuracy, appropriateness and empathy ^{(International Journal of Medical Informatics 2025)}.

○ What holds

The moat here is accountability backed by licensure, and it is unusually explicit. GMC guidance states that doctors remain responsible for decisions made using AI, which means software can draft and suggest but cannot hold the liability ^{(General Medical Council n.d.)}. Deployment is gated by institutions rather than by capability, with NHS England issuing conditions for AI-enabled ambient scribing products in April 2025 ^{(NHS England 2025)}. The applicability research separates the augmented parts of the role from the resistant ones, and physical examination sits firmly on the resistant side ^{(Microsoft Research 2025)}. The study that beat physicians on written questions did not test examination, longitudinal care or statutory sign-off at all ^{(International Journal of Medical Informatics 2025)}.: nothing in it touches the consulting room.'

Which tasks go first

Split the job by task and the picture resolves. Note taking, letter drafting and administrative correspondence are already being absorbed at scale, with saved documentation hours and reduced screen time measured across millions of encounters ^{(Kaiser Permanente Division of Research 2025)}. Answering common clinical questions from text is a task where current models perform strongly against family physicians in controlled comparison ^{(International Journal of Medical Informatics 2025)}. Clinical decision support is real but minority use, reported by 28% of GPs who use AI at all, and it functions as an assistive second opinion rather than an autonomous one ^{(Nuffield Trust and Royal College of General Practitioners 2025)}. What stays human is the physical examination, the continuity of a patient relationship over years, and the signature on the treatment decision, since responsibility for AI-assisted decisions remains with the doctor ^{(General Medical Council n.d.)}. The applicability evidence frames the whole role as augmented rather than substituted, and the group pattern across healthcare is the same: embodiment plus in-person care plus licensure plus life-safety accountability ^{(Microsoft Research 2025)}.

What to do about it

Become the person who edits the machine's draft well

Ambient documentation is arriving whether or not any individual clinician asks for it, with one enterprise rollout targeting more than 12,000 clinicians by 2026 ^{(Abridge 2025)}. Because responsibility for decisions made using AI stays with the doctor, the skill that gains value is fast, sceptical review of generated notes and suggestions ^{(General Medical Council n.d.)}. That is a learnable habit, and it is the part of the workflow the tool cannot own.'

Learn your institution's authorisation route, not just the tool

NHS England's April 2025 guidance makes ambient scribing adoption conditional on institutional authorisation rather than on what the product can technically do ^{(NHS England 2025)}. Clinicians who understand the governance path tend to be the ones consulted when a practice or trust chooses a system. Only 28% of surveyed GPs reported using AI at all, so the pool of people with practical deployment experience is still small ^{(Nuffield Trust and Royal College of General Practitioners 2025)}.'

Weight your development toward the untested half of the job

The comparison that favoured a model over family physicians covered written clinical questions and explicitly did not cover physical examination, longitudinal care, statutory sign-off or liability ^{(International Journal of Medical Informatics 2025)}. Examination skill, continuity of care and complex-case judgment are precisely where the applicability evidence says the role resists ^{(Microsoft Research 2025)}. Depth there compounds, and it is not the kind of depth a documentation tool erodes.'

Tools and courses that fit this situation

Below are paid tools and courses that fit a role where the exposure is concentrated in documentation and written clinical reasoning. They are general AI-literacy resources rather than clinical systems, and none of them changes the exposure itself.

Some links above are affiliate links. If you buy through them, Jobs After AI may earn a commission at no extra cost to you. It does not change which tools appear here.

Common questions

How is the 76 out of 100 score for physicians worked out?

It weighs strong augmentation of diagnosis and documentation against the moat formed by licensure, physical examination and treatment accountability ^{(Microsoft Research 2025)}. Confidence in this reading is medium, so treat 76 and the rank of 128th out of 159 as directional rather than precise. The evidence ledger below the page carries every claim behind it, including the ones that cut the other way.'

Is this about physicians losing jobs, or about their tasks changing?

Tasks, on the current evidence. The largest documented deployment saved nearly 16,000 documentation hours across 7,260 physicians and reported no change in physician headcount ^{(Kaiser Permanente Division of Research 2025)}, and a 12,000-clinician enterprise rollout was announced with no reduction, layoffs or lowered hiring target attached ^{(Abridge 2025)}. Official employment in the closest BLS category rose from 98,590 in May 2020 to 112,010 in May 2023, with BLS attributing none of it to AI ^{(US Bureau of Labor Statistics 2020)}.'

Which physician tasks are still going to AI?

Documentation first and most decisively, used by 57% of GPs who use AI, followed by administrative tasks at 44% ^{(Nuffield Trust and Royal College of General Practitioners 2025)}. Text-based clinical question answering is also strong, with one 2025 study rating a model above family physicians on accuracy, appropriateness and empathy ^{(International Journal of Medical Informatics 2025)}. Clinical decision support is used by 28% of AI-using GPs and functions as an assistive input ^{(Nuffield Trust and Royal College of General Practitioners 2025)}.'

What would have to change for this score to fall?

Two things, and neither is purely technical. Professional accountability would have to move off the physician, whereas current GMC guidance keeps doctors responsible for decisions made using AI ^{(General Medical Council n.d.)}, and institutional gatekeeping would have to loosen, whereas NHS England's 2025 guidance ties adoption to authorisation ^{(NHS England 2025)}. Workforce projections currently run the other way, with the AAMC forecasting a shortfall of up to 86,000 physicians by 2036 ^{(Association of American Medical Colleges 2024)}.'

Related occupations

The evidence behind this page

How this score was built

The AI-Proof Score is an estimate of how exposed an occupation's core tasks are to AI over roughly the next three to five years, on a 0–100 scale where higher is safer. It measures the exposure of tasks, which sits upstream of, and is softer than, job loss — it is not a probability that this job disappears, and not a claim about any individual doing it. The full derivation, the four research families behind it and its limitations are set out in the methodology, and every occupation is ranked in the AI Job Impact Report 2026.