In 2016, Geoffrey Hinton said we should stop training radiologists. "It's just completely obvious that within five years, deep learning is going to do better than radiologists."
It is 2026 now. Ten years later. Radiologists were not replaced. Demand for their work went up, and there is still a shortage.
This is one of the most instructive AI predictions we have, because the technology did not fail. The economics simply behaved differently than expected. If you lead a team of experts and someone tells you AI will make their jobs disappear, this story is worth ten minutes.
The Prediction Was Right About the Technology
It is tempting to read the radiology story as "the AI did not work". That is not what happened. According to the FDA's list of AI-enabled medical devices, there are over 1,000 FDA-cleared AI tools for radiology. The technology works. Hinton was not wrong about the capability.
Where the prediction went wrong was the step from "a machine can do part of this job" to "we will need fewer people doing this job". That step assumes the amount of work stays the same. It rarely does.
Here is what happened to radiologists instead:
Demand for imaging increased, not decreased.
There is a global shortage of radiologists, not a surplus.
AI takes over routine scans, so radiologists can focus on the complex cases.
The Jevons Paradox, Explained
AI made imaging cheaper and faster. Cheaper, faster imaging meant doctors ordered more scans. More scans meant more work for radiologists, not less.
Economists have a name for this. The Jevons paradox: when technology makes a resource more efficient to use, total consumption of that resource often goes up, not down. The English economist William Stanley Jevons described it in 1865 in The Coal Question, after noticing that more efficient steam engines did not reduce Britain's coal use. They increased it, because coal power suddenly paid off in places where it had been too expensive before.
Efficiency lowers the cost of each scan. Lower cost means more scans get ordered. More scans mean more work for the experts who read them.
The pattern repeats wherever something valuable becomes cheaper:
Coal got more efficient in the 1800s. We used more coal, not less.
Computing got cheaper. We did not end up using fewer computers.
AI makes diagnosis faster. We do not need fewer diagnosticians.
Every time we automate something, we discover uses for it that we could not afford before. The work that was rationed by cost starts to get done.
Why This Matters Beyond Radiology
The same confident predictions are being made right now about software engineers, designers, writers and analysts. "AI will replace X within Y years."
Maybe. But we have heard that before, and so far the track record of these predictions is poor. They tend to count the tasks a machine can take over and forget to ask what happens to demand once those tasks get cheap.
My own view is that software is the next obvious case. Building software has become dramatically cheaper. I do not think that means we will build the same amount of software with fewer people. I think we will build far more software, for problems that were never worth a development project before.
That is an opinion, not a forecast with a number attached. But it is the question worth asking in any field: when this work gets cheaper, will people want the same amount of it, or much more?
What This Means for Leaders of Expert Teams
If you run a team of specialists, in healthcare software and services or in any business built on expert judgment, the radiology story suggests a different set of questions than "how many people can we save?".
Plan for more demand, not fewer people
When a service gets faster and cheaper to deliver, customers usually ask for more of it. Assume your experts will be busier, and decide in advance where that extra capacity should go.
Point AI at the routine work
The radiology pattern is AI taking over the routine cases so specialists spend their time where judgment matters. Look for the recurring, well-described part of your team's work first.
Redesign the workflow, not just the task
More volume puts pressure on everything around the task: review, handovers, scheduling, documentation. If only the task gets faster, the bottleneck simply moves to the next step.
Keep a person responsible for the hard cases
Decide who reviews what the AI hands over, and how complex cases reach the right expert. That design is what turns cheaper work into better work.
Routine cases flow through with AI support. Complex cases reach the expert, whose time now goes where it matters most.
This is the kind of work we do with expert teams: one real workflow, redesigned together. You can read more about how we work .
Pay Attention to the Jevons Paradox
The most confident AI prediction of the last decade got the technology right and the economics wrong. That is worth remembering the next time someone tells you exactly which jobs will disappear and when.
The better question for your team is not whether AI can do part of the work. It is what becomes possible once that part gets cheap, and whether your people are ready to meet the demand that follows.
Frequently Asked Questions
- Will AI replace radiologists?
- So far it has not. Geoffrey Hinton predicted in 2016 that deep learning would outperform radiologists within five years. Ten years later AI tools for radiology are widely cleared and used, yet demand for radiologists has grown and a shortage persists. AI takes over routine scans while radiologists focus on complex cases.
- What is the Jevons paradox?
- The observation, first described by William Stanley Jevons in 1865, that making a resource more efficient to use often increases its total consumption. More efficient steam engines led Britain to use more coal, not less, because coal became worth using in more places.
- Does the Jevons paradox apply to AI and jobs?
- It can. When AI makes a service faster and cheaper, people often ask for more of it, which can increase the work for the experts involved. It is not a guarantee for every role, but it is a strong reason to question predictions that only count the tasks a machine can take over.
- How should healthcare organisations plan for AI?
- Plan for higher demand rather than fewer people. Use AI for routine, well-described work, redesign the workflow around it, and keep a named expert responsible for complex cases and for reviewing what the AI produces.
How we work · Work · Insights