Will AI replace radiologists? The workforce data says no
The UK is 32% short of radiologists, ten years after AI

The UK is now 32% short of the radiologists it needs. That's ten years after we were told to stop training them.
So the "AI will replace radiologists" story has a slight problem. AI in radiology is real and useful. The jobs data just points the other way (for now, and probably for a while).
The consensus: AI will replace radiologists
In 2016, Geoffrey Hinton said people should stop training radiologists, because deep learning would beat them within five years (The Decoder, reporting the New York Times).
The logic was neat:
→ radiology is reading images → AI is very good at reading images → so radiologists are first in line
It stuck. It still gets repeated to medical students every year.
What the radiology workforce data actually says
1. The UK shortage is getting worse, not better. The Royal College of Radiologists' census, out on 18 June:
→ 32% shortfall, over 2,300 radiologists, up from 29% → on track for 40% (over 4,050) by 2030 without action → demand for scans growing more than twice as fast as the number of radiologists → 39% of radiologists set to retire within 10 years
2. The NHS is paying a fortune to plug the gap. RCR figures from May:
→ £362 million on outsourcing, overtime and locums in 2025 alone → £1.4 billion over five years → 940,900 scan results in England took over a month in 2025
3. US training posts hit a record. The 2026 Match offered 1,241 diagnostic radiology positions, up from 1,210 last year and 1,132 in 2022 (RT Medical, reporting NRMP data).
4. The flagship AI adopter hired MORE radiologists. Mayo Clinic went from about 260 radiologists in 2016 to more than 400, up 55%, whilst running over 250 AI models (The Decoder).
Even Hinton now says he spoke too broadly and got the timing wrong.
Where the consensus is right: AI is changing radiology work
It's not wrong that AI changes the job. It's wrong about what that looks like.
The best evidence is the Swedish MASAI trial, with around 106,000 women (Lund University, January):
- AI sorted each mammogram into single or double reading by risk score
- it found 29% more cancers than standard screening
- radiologists' screen-reading workload fell 44%
- missed cancers between screenings: 82 with AI vs 93 without
Notice who still reads the scans. Radiologists. AI decided how many of them needed to look.
bad headline: "AI replaces radiologists" accurate headline: "AI lets a short-staffed department read more scans, better"
And here's the bit that worries me. US diagnostic radiology applicants fell to 1,741, about 14% below the 2023 peak of 2,014 (RT Medical). If the replacement story scares good students off, it makes the shortage worse.
What med school won't teach you about AI in radiology
Medical training is brilliant on anatomy and pathology. It's thin on:
- how AI tools are validated. Sensitivity, specificity, the population it was trained on
- workflow. MASAI's gain came from changing who reads what, not from a smarter algorithm alone
- reading jobs data. One quote from 2016 shaped a decade of career advice
We covered safe ways to study with these tools in AI tools for medical students.
The critical-thinking move: 4 questions for any "AI will replace X" claim
- What does the job actually involve? Reading images is one part. Talking cases through with other doctors and owning the final report are others
- Is demand fixed? Scan demand is growing twice as fast as the workforce. Tools that speed things up get absorbed, not banked as job cuts
- What does the headcount data say? Prediction vs census. Census wins
- Who still signs the report? If a human is accountable, the human stays
Same questions work outside medicine. Software has the same debate, and in AI agents and the junior developer job market the data is messier, but the "who reviews it" point is identical.
What to do this month
- Read the RCR census summary and pull three numbers you'd use in an interview
- Read the MASAI results and explain, in two lines, why workload fell
- Shadow a reporting session if you're on placement. Ask how AI flags or triage tools fit in, and who checks them
- Look up one AI tool used in your local trust and find its validation study. If you can't find one, that's an answer too
More on AI across medicine is on our healthcare hub.
so TLDR: ten years on, the UK radiologist shortage is getting worse and the US is offering record training posts. AI is changing how scans get read, not who's accountable for them.
If you were choosing a specialty today, would the "AI will replace it" headlines put you off radiology, or make it look more interesting?
Sources
- 'Alarm bells should be ringing': staff shortages causing 'extremely dangerous' delays to diagnosis and cancer care, doctors warnThe Royal College of Radiologists · 18 JUN 2026
- NHS spent record £1.4bn on short-term fixes to radiologist shortage in past 5 yearsThe Royal College of Radiologists · 26 MAY 2026
- Match Day 2026: Record Positions, Fewer ApplicantsRT Medical Systems (reporting Radiology Business and NRMP data) · 23 MAR 2026
- AI support in breast cancer screening: Fewer missed cancer casesLund University · 30 JAN 2026
- Geoffrey Hinton's wildly overconfident AI prediction failed, now it's a lesson in humilityThe Decoder (reporting The New York Times) · 15 MAY 2025
Get the next issue by email
theVANTAGEd is written by students to bring critical thinking to AI across real industries. One issue every two weeks.
