Radiologist
In 2016 the prediction was: no more radiologists by 2025. Reality: more radiologists than ever, with AI as a co-reader. This is the template for understanding the current cycle.
How the work changes over time
Tasks
Reading routine imaging (chest X-rays, screening mammograms)Augmented
High-volume, well-characterized studies with clear normal and abnormal patterns.
Why: Models are already at or above human level for many narrow tasks. The bottleneck is regulatory approval and workflow integration, not capability.
Reading complex cross-sectional imaging (CT, MRI)Assisted
Studies with many findings, often correlated with clinical history and prior imaging.
Why: AI flags findings well; integrating them into a clinically meaningful report is harder. That gap is closing.
Interventional procedures (biopsies, drainages, ablations)Assisted
Image-guided procedures where the radiologist is operating, not just reading.
Why: Robotics help, but the radiologist is still in the room making decisions about a specific patient's anatomy in real time.
Discussing findings with referring cliniciansUntouched
The five-minute phone call that often does more than the formal report.
Why: Clinical reasoning shared between two specialists about a specific patient. Not what models do well.
Communicating significant findings to patientsUntouched
Increasingly, radiologists deliver results directly — especially for screening and incidental findings.
Why: A person telling another person what their scan shows. Hard to displace and arguably should not be.
Quality assurance and case review for the departmentAssisted
Catching errors, teaching from cases, improving the department.
Why: AI is excellent at catching patterns in error data; humans still decide what to do about it.
Reading studies the AI flags as uncertainAssisted
A growing category: the model says "I am not sure," and a human reads.
Why: This is the work that survives — the hard cases the model cannot confidently call.
Subspecialty consults (neuroradiology, pediatric, musculoskeletal)Assisted
Deep expertise in a narrow domain.
Why: Subspecialty AI is harder to train because data is scarcer. The humans here have more runway.
Teaching residents and fellowsUntouched
Training the next generation of radiologists.
Why: Live teaching of clinical judgment is human. AI tools become part of what residents learn to use.
Medical-legal review and expert witness workUntouched
Forensic work where liability is the whole point.
Why: A model does not testify. This corner of the practice is robust.
The honest take
Ten years ago, a famous AI researcher said radiologists should "stop training, they will all be out of work in five years." There are now more radiologists than there were then. Demand outpaced what AI could absorb. Imaging volumes grew. Subspecialties multiplied. The prediction was not wrong about the technology — narrow image classifiers really did get very good. It was wrong about how work actually changes.
You are now reading studies alongside an AI that is, for many tasks, as good as you. That is the actual experience of the job in 2026. The model finds the obvious fracture; you confirm it, dictate the report, and move on. The model is uncertain about a complex MRI; you read it carefully and make the call. The interesting and difficult cases come to you faster because the routine work is faster. Your throughput is up. Your error rate is down. Your job is more demanding, not less, because the easy reads no longer give you a break.
Where the headcount question lands in fifteen years is genuinely uncertain. It depends on imaging volume growth, regulatory tolerance for autonomous reads, and how much of the work shifts to referring clinicians using AI directly. The role is changing more than disappearing.
What protects this role
- Medical licensing as a regulatory gate.
- Liability for clinical decisions.
- Clinical judgment in non-routine cases.
- The legal and ethical requirement of a physician in the loop, in most jurisdictions.
What to do Monday
- Use the AI tools your department has deployed. Radiologists who supervise the model well are noticeably more productive than those who fight it.
- Consider subspecialty training if you have not already. The narrower the domain, the slower the AI catches up.
- Spend time on the work that the model cannot do: interventional procedures, complex consults, direct patient communication.
- Pay attention to how your reports are being used — by AI systems, by referring clinicians, by patient portals. The downstream is changing fast.