LO
Radiologist
Reads images to answer a clinical question — and signs for the answer, including for what nobody thought to ask about.
This is not a probability of losing your job. It combines how much of the role's task load is exposed to automation with how far adoption has actually gone — useful for comparing occupations on one consistent basis, and for nothing else.
Written for diagnostic radiologists reading studies in a hospital or teleradiology setting. Interventional radiology is included as one task but is really a different job with a different exposure. Radiographers and technologists — the people who acquire the images — are a separate occupation this site does not cover yet, and their picture is not the same.
The evidence base holds verified records for other occupations, but not one for this one yet. Until it does, the analysis below is reasoning about task structure and known technical capability — for this job in particular it is not backed by traceable sources, and we would rather say so than cite things we have not verified. An empty section here is a gap in our coverage, not a finding about the work.
What is actually changing#
The unit of analysis is the task, not the job title. A role is not replaced — its task mix shifts.
Is this your job? Say so and this page narrows to your share of it.
A job title is a bundle of tasks bought together, and no two people hold the same bundle. Nothing is sent anywhere — it stays in this browser.
Reading the routine study
Automating≈ Platform inferenceHigh-volume, pattern-clear work: screening mammograms, chest films, follow-up scans where the question is narrow and the answer is usually normal.
This task has the four properties that make anything a strong case for machine work: the input is an image, the output is a label, correctness is checkable against a later outcome, and the training data exists in enormous quantity because every study was already read and reported by someone. It is also the task the 2016 prediction was actually about, and on this task the prediction was directionally right even though its timeline and its conclusion were not.
Says nothing about how much of a radiologist's day this is, and the share differs enormously between a screening programme and a tertiary hospital. It also says nothing about deployment: a regulator clearing a device for market is not a hospital using it, and this site holds no verified record of either for this occupation yet. What would settle it: a health system's own report of what share of studies are machine-read before a person sees them.
Catching what nobody asked about
Still human-led≈ Platform inferenceThe scan was ordered for a back injury; the thing that matters is in the corner of the image and has nothing to do with the back.
A detection model is trained on the question that was asked. The incidental finding is by definition the answer to a question nobody asked, and it is a large part of what makes reading valuable rather than merely accurate. Improving a model on its own task does not move this, because it is not that task.
A judgement about the structure of the work, not a measurement. We hold no record quantifying how often incidental findings change management, and the answer differs by modality and by population. It also does not claim people are good at this — missed incidentals are a known and studied failure.
Deciding whether to image at all
Being augmented≈ Platform inferenceProtocolling: whether this study answers the clinical question, which sequence, how much contrast, how much dose — and whether to say no.
Decision-support tools genuinely help here: appropriateness criteria are rule-shaped and a system can apply them consistently where a busy person will not. What does not transfer is the refusal — declining a study a referring clinician wants is a judgement with a cost that lands on a named person, and the value of the radiologist in that moment is precisely that they can be argued with.
Nothing here says how often radiologists actually protocol studies rather than rubber-stamping them, and in many settings that step has already been delegated or automated away for reasons unrelated to AI.
The report somebody acts on
Being augmented≈ Platform inferenceTurning what you saw into words a clinician will act on — including how much to hedge, and whether to recommend a follow-up that costs money and worry.
Drafting and structuring reports is well-suited to generation and is where most of the visible time saving is claimed. The part that does not move is the calibration: a sentence that hedges too much is useless and one that hedges too little causes an unnecessary operation, and where that line sits depends on this patient, this referrer and this health system.
No verified record on this site measures report-drafting time saved, and vendor claims about it are marketing from a party with a stake in the answer. What would settle it: a health system publishing its own before-and-after turnaround times.
Procedures done inside a person
Still human-led≈ Platform inferenceImage-guided biopsies, drains, embolisation — interventional work where the image is the guidance rather than the product.
Physical, irreversible, performed on a person who is present and can be harmed. Robotics in this area assists a hand rather than replacing one, and regulation attaches the act to a licensed individual. This is the half of radiology that the 2016 prediction did not address at all.
Interventional radiology is really a separate occupation with its own training path, and treating it as one task of this page understates how different it is. It also says nothing about volume, which is where the actual pressure on this work comes from.
Answering for the machine that read it
New task≈ Platform inferenceValidating a tool on your own population before switching it on, watching for drift after, and being the named person on a report the machine helped write.
Work that did not exist before these tools entered the reading room, and it lands here because the signature does. A model validated elsewhere is not validated here — performance moves with scanner, protocol and population — so somebody local has to check, and checking is a radiologist's skill applied to a machine instead of to a patient.
New work appearing is not new headcount, and in most departments this is absorbed by whoever is already there. We hold no record of a hospital staffing it deliberately.
Which technologies matter here#
Four separate signals. They are deliberately not added together — a job exposed to two technologies is not twice as exposed.
How it got here#
The index is not a static number. This is where it would have sat at each capability checkpoint since ChatGPT — reconstructed, and labelled as such.
A high starting point that has nothing to do with generative AI: computer-aided detection was cleared for mammography in the late 1990s and deployed widely, so this occupation entered the period already partly automated — and already knowing how that felt. The climb through 2023-2025 is narrow detectors and triage tools reaching the reading room in quantity. It flattens well below its most exposed task, and the reason is what the famous 2016 prediction left out: a clearance is not a deployment, a detection is not a signature, and the incidental finding is the answer to a question the model was never asked. Note what the curve cannot show: whether cheaper reading raises the number of studies ordered. That is the open question for this occupation, and it points the other way from the curve.
A flat line is not a forecast of safety. It says which tasks automation has reached so far — the occupations that moved least here are the ones where the constraint is physical or regulatory, and both of those can change.
Recent changes#
No verified events recorded yet.
This section will fill from the monitoring pipeline as events are collected, de-duplicated, graded and linked to the tasks above. An empty list here means we have not verified anything — it does not mean nothing is happening.
"We found no news" is not the same as "you are safe."
What this means for you#
You are choosing a specialty that a very famous person publicly advised nobody to enter. Worth knowing why that advice did not hold: it was about one task, and the job turned out to be several — including signing for a miss, seeing what nobody asked about, and doing procedures. It is also worth knowing the honest risk: if cheap reading makes imaging cheaper, volume rises, and a job defended by volume is defended by something that can change.
The prediction being wrong is not the finding; the reason it was wrong is. It measured a task and concluded about a job. Apply the same test to your own week: the hours spent on studies where the question is narrow and the answer is usually normal are the exposed hours, and they are also the ones most easily counted by an administrator building a business case.
Your options#
Four directions, each with its real constraints and one thing you can test this week. Continuing as you are is a legitimate choice — it just has to be a chosen one.
Be the one who validates the tools
A model validated elsewhere is not validated here, and somebody local has to sign that it works on this population, on these scanners. Most departments have not named that person.
It needs statistics you were probably not taught, and it is unglamorous work that competes with reading volume you are measured on.
Find one AI tool running in your department and ask what population it was validated on. If nobody can tell you, that is the job.
Move toward the procedure
Interventional work is physical, irreversible and licensed to a person — the combination this site's evidence consistently finds hardest to displace.
It is a different training path and a different life: call, radiation exposure, and a patient in front of you rather than a worklist.
Spend a session in the IR suite and count how much of what happens there could be specified in advance. Very little of it can.
Go where the images are used, not read
Clinical informatics, imaging programme design and screening policy all need someone who understands what a reading actually establishes — and that is a scarce kind of person in rooms where those decisions get made.
You stop being the person who reads, and the skill you traded on decays faster than you expect.
Look at one screening programme's own published outcome data and check whether its recall rate is where the protocol says it should be.
Common questions#
This occupation is the reason to distrust that question. A decade ago a very well-known researcher publicly advised that training should stop, on the grounds that the machine would be better within five years. On the narrow task he named, the machines did get very good. The job did not go away, because the job was several things and he had measured one. A signal you can check yourself, and a better use of the worry: of your last week, how many hours went to studies where the question was narrow and the answer was usually normal. That number is the exposed part, and it is also the number an administrator can count.
Wrong in its conclusion, and worth understanding rather than dismissing, because the mistake is the one everybody makes. It took a task — reading a high-volume, pattern-clear study — and concluded about a job. On that task it was directionally right. What it left out was who signs for a miss, who sees the finding nobody asked about, who declines the study that should not be done, and whether cheaper reading makes demand rise rather than fall. Dismissing it as hype teaches nothing; reading it as a category error tells you exactly what to check about your own work.
A clearance says a product may be sold; it does not say a hospital bought it, switched it on, or kept it on. This site keeps those three things separate on purpose, and the gap between them is where most predictions about this field have gone wrong. It is also worth knowing which direction the products point: most are narrow detectors and triage tools that change the order of a worklist, which changes when a person looks rather than whether one does.
This site does not tell anyone which specialty to choose, and a page that did would be worth less. What it can give you is the shape of the bet: the exposed half is the high-volume narrow read, the durable half is accountability, incidental findings and procedures, and the open question is volume — whether cheaper imaging means more of it. Ask that question of the people already doing it, and notice whether they answer with what their department actually measures or with what they hope.
Method and sources#
- Assessment date
- 2026-09-12
- Basis of the task judgements
- 0 evidence-backed · 6 platform inference · 0 not enough evidence
- Verified events
- 0