You gave years and a fortune to medicine because no machine could do it. Now AI reads scans and writes notes. What that means for your career.
This in-depth guide covers everything you need to know about will ai replace doctors? radiology, scribes and the data. Based on verified income data and real-world case studies from our database of 138 side hustle tactics.
Mostly no, for now. The world is short of doctors and official forecasts still expect physician jobs to grow, yet AI is already taking over pieces of the job that doctors spent years learning: reading scans, answering exam-style clinical questions, and writing up the visit. The people most exposed right now are the ones whose work is mainly documentation, and the specialists whose value was a trained eye on an image are watching the machines close in.
You worked harder than almost anyone you know to get here. The exams, the night shifts, the debt that is larger than some people's mortgages, the birthdays you missed. Your parents told the whole family when you got in. You chose medicine partly because it was supposed to be the one career no machine could take, and partly because you wanted to be the person who helps.
Then a headline says a chatbot beat doctors at diagnosis. A colleague jokes that the AI scribe writes better notes than the residents. And at 2am, scrolling after a long shift, you wonder whether the next decade of your career is going to be spent signing off on a computer's opinion.
What AI can already do in medicine
Start with the exams, because that is where the headlines started. MedQA is a benchmark built from US medical licensing exam style questions. The Stanford AI Index 2025 reported that OpenAI's o1 set a new record of 96.0% on MedQA, 5.8 points above the best score from 2023, and warned the benchmark may be nearing saturation. A test that was meant to sort humans by clinical knowledge no longer separates the top models from each other.
Diagnosis studies are more unsettling. In a randomised trial published in JAMA Network Open in October 2024, 50 US physicians worked through complex cases with or without GPT-4. Doctors with the AI scored a median 76% on diagnostic reasoning and doctors without it scored 74%, while GPT-4 working alone scored 92%. Giving doctors the tool barely helped. The tool on its own did better than both groups.
In June 2025, Microsoft published results for its MAI Diagnostic Orchestrator on 304 hard case records from the New England Journal of Medicine. Its best setup correctly solved 85.5% of cases, against a mean of 20% for 21 practising physicians with 5 to 20 years of experience. Microsoft notes the caveats itself: the work was not yet peer reviewed, it is not approved for clinical use, and the physicians were barred from using colleagues, textbooks or AI, which is not how real medicine works.
A question to sit with: if a model can match you on the hardest written cases, what part of your day does a patient actually come to you for?
The FDA list keeps growing, and it is mostly radiology
If you want one number that shows where AI has landed in medicine, look at the US Food and Drug Administration's public list of AI-enabled devices. As of its September 2026 update, which covers decisions through June 29, 2026, the list contains 1,614 authorised devices, and 1,230 of them, about 76%, sit under radiology. The pace is speeding up: the list records 235 devices for 2024, 335 for 2025, and 181 in just the first half of 2026.
Cardiology is a distant second at 154 devices. Everything else trails far behind. That is why the question "will AI replace radiologists" gets asked so often. Radiology is the specialty where software has been cleared to do the most.
The prediction that made radiologists nervous
In 2016, at a machine learning conference, Geoffrey Hinton, who went on to share the 2024 Nobel Prize in Physics, said: "People should stop training radiologists now." As he later described it himself, the forecast was that neural networks would beat radiologists at reading scans within five years. Every medical student deciding on a specialty heard about it.
Here is what happened since. Radiology did not shrink. Works in Progress reported in September 2025 that US radiology residency programmes offered a record 1,208 positions in 2025, up 4% on 2024, and that average radiologist income reached $520,000, more than 48% above the 2015 average and the second-highest of any specialty. The same piece points to a 2012 study that found staff radiologists spent only 36% of their time directly interpreting images. The rest was consulting, procedures, teaching and talking to colleagues and patients.
Hinton has since revisited his own words. In a December 2023 interview with cardiologist Eric Topol, he said: "So in 2016, I made a daring and incorrect prediction." He explained that he had meant it "for interpreting medical scans" rather than for everything a radiologist does, and conceded: "I was wrong about that." He did not give up the direction, though: "I think I was off by about a factor of three, but I'm still convinced I was completely right in the long term." He predicted that in 10 years AI would "routinely" give a second opinion on scans, and that in about 15 years "the doctor's opinion will be the second one."
When The New York Times revisited the prediction in May 2025, Hinton said he expected most image interpretation to be done by a "combination of A.I. and a radiologist," making radiologists "a whole lot more efficient".
So radiologists were not replaced. But read Hinton's timeline again. He is saying the reversal, where the machine reads first and the human checks, is roughly one career length away. A radiologist who starts residency this year will be mid-career when he expects it to happen.
Think about what that means in a real life. You finish training in your early thirties with a large loan. You buy a home near the hospital and plan your children's schooling around a salary that assumes your skill stays scarce. Then, somewhere around year ten of practice, the hospital starts asking why it needs two radiologists to check a machine's report when one would do. Nobody fires you. They simply stop replacing the colleagues who retire, and the overnight reads go to a smaller team with better software. That is the pattern Stanford economists found when they studied young workers in AI-exposed jobs: the decline came mainly through reduced hiring rather than people being let go.
The scribe in the room
The part of medicine AI has changed fastest is the part doctors hated most: paperwork. Ambient AI scribes listen to the consultation and draft the clinical note.
The American Medical Association's survey, released on March 12, 2026, found that 81% of physicians now use AI in their practice, up from 38% in 2023, with documenting clinical care among the most common uses. The same survey found 88% of physicians worried about losing skills, most of all those with ten years or less in practice.
The largest published rollout is at The Permanente Medical Group in California. Over 63 weeks from October 2023 to December 2024, 7,260 physicians used AI scribes across 2,576,627 patient encounters, saving an estimated 15,791 hours of documentation, the equivalent of 1,794 eight-hour workdays.
For doctors that is relief. For the people whose job was the note, it is a different story. The US Bureau of Labor Statistics projects that medical transcriptionist jobs will fall 4% from 2025 to 2035, and it names the cause: advances in speech recognition and natural language processing that let physicians document visits in real time. For medical coders and records specialists, BLS still projects growth of 8%, but it warns that the adoption of AI-powered solutions may reduce demand by making coding more efficient.
Nurses, admin staff and the phone calls
Nursing is the least replaceable work in the building, because so much of it is physical and human. BLS projects registered nurse employment to grow 6% from 2025 to 2035, with about 180,800 openings a year.
What AI is going after is the nurse's phone time: discharge follow-ups, appointment reminders, intake questions, screening outreach. Hippocratic AI, one company building voice agents for this work, says on its website that its agents have handled more than 250 million patient interactions, with more than 31,000 patients escalated to human nursing care. The company says its agents do not diagnose or prescribe. Those calls used to be someone's shift.
Front-desk work is mixed. BLS projects medical secretaries and administrative assistants to grow 5% over the decade, while secretaries and assistants overall decline 2%. Healthcare is still hiring people to run the front of the house. The question is how long scheduling and intake stay human when an agent can do them at 2am without a break.
A question to sit with: which parts of your week are spent on work a voice agent or scribe could draft, and which parts only happen because you are in the room?
What the official forecasts say
The official forecasts are not frightening for doctors. BLS projects physician and surgeon employment to grow 4% from 2025 to 2035, from about 862,800 jobs, with about 22,100 openings a year and a median wage of $275,930 in May 2025.
The bigger problem is that there are not enough doctors. The Association of American Medical Colleges projects a US shortage of up to 86,000 physicians by 2036. Globally, the World Health Organization projects a shortfall of about 11 million health workers by 2030, mostly in low and lower-middle income countries.
That is the honest counter-evidence, and it is strong. In a world short of doctors, AI that makes each doctor faster tends to mean more patients seen rather than fewer doctors employed, at least for the next decade.
Where forecasts disagree is on the longer run. BLS projections assume today's way of working continues. Hinton's 15-year view, and the speed of the FDA list, assume it will not. The IMF's Kristalina Georgieva wrote in January 2024 that about 60% of jobs in advanced economies "may be impacted by AI". Medicine is full of exactly the kind of knowledge work those estimates describe.
The numbers in one place
| Measure | Figure | Source |
|---|
| Top AI score on MedQA (licensing-style questions) | 96.0% (o1, 2024) | Stanford AI Index 2025 |
| Diagnostic reasoning: GPT-4 alone vs doctors | 92% vs 74% (no AI) and 76% (with AI) | JAMA Network Open, Oct 2024 |
| Microsoft MAI-DxO on 304 NEJM cases | 85.5% vs 20% for 21 physicians | Microsoft AI, June 2025 |
| Members of the public using chatbots to identify conditions | Under 34.5% (models alone: 94.9%) | Oxford, Nature Medicine 2026 |
| FDA AI-enabled devices | 1,614 total, 1,230 in radiology (76%) | FDA list, through June 2026 |
| Physicians using AI | 81% (2026), up from 38% (2023) | AMA |
| AI scribe rollout, Permanente Medical Group | 7,260 doctors, 2.58 million encounters, 15,791 hours saved | AMA |
| US radiology residency positions | Record 1,208 (2025) | Works in Progress |
| Physicians and surgeons, BLS projection | +4% (2025 to 2035) | BLS |
| Medical transcriptionists, BLS projection | -4% (2025 to 2035) | BLS |
| Projected US physician shortage | Up to 86,000 by 2036 | AAMC |
Why the machine still needs you, for now
The study that should reassure doctors most came out of Oxford. In a randomised study of 1,298 UK participants published in Nature Medicine in 2026, language models on their own identified the relevant conditions in 94.9% of scenarios, but when members of the public used those same models, they identified conditions in under 34.5% of cases and did no better than people using ordinary web searches.
The knowledge is in the model. Getting the right story out of a frightened person, knowing which detail matters, and deciding what to do next is still the clinician's work. That is the gap your training fills.
But notice what the problem has become. Ten years ago, a doctor's value was knowing things few others knew. Today the model knows them too, and your value is the judgement, the examination, the conversation and the accountability. Those are real. They are also harder to put on a CV, and harder to defend in a budget meeting.
A question to sit with: if your hospital measured you only on what a model cannot do, how would your job description read?
What this means for you this month
If you are a doctor, the lowest-risk move is to get fluent with the tools before your employer makes the choice for you. Use the scribe, check its notes line by line, and keep a record of what it misses. Clinicians who can judge AI output are the ones hospitals will want on the committees that decide how it is used. The 88% of physicians worried about losing skills are right to protect them: keep doing the reasoning yourself before you read the machine's answer.
If you are a medical student, radiology and every other specialty are still hiring, and the shortage is real. Choose the specialty you want, then aim for the parts of it that need hands, conversations and procedures.
If you work in transcription, coding or scheduling, the pressure is nearest to you. Clinical knowledge has a market outside the hospital: some people teach it through online tutoring or build courses for students and nurses, and others are paid to grade medical answers in AI training jobs. If you are thinking about a bigger move, read career change at 40 and see how other professions are faring in will AI replace lawyers and jobs AI will replace.
You trained to be the person in the room. For now the room still needs you, and the work you do there is changing underneath you.