Radiology, ten years after “radiologists are finished”

In 2016, Geoffrey Hinton — as close to a founding father as deep learning has — told an audience that we should stop training radiologists, because it was, in his view, obvious that image-recognition systems would outperform them within a handful of years. It was a reasonable prediction from the person best positioned to make it, about the medical specialty most exposed to exactly what neural networks do best.

A decade later, radiology residency positions still fill. Imaging volumes have grown. Health systems in the US and UK have spent recent years describing radiologist shortages, not surpluses. And — this is the part that should reorganize how you read every confident displacement prediction — the AI worked. Hundreds of imaging algorithms have been cleared by regulators. Models genuinely do read certain scans at or above specialist level in controlled settings. The technology succeeded and the prediction failed anyway, which makes radiology the single best control case we have for the claim that follows every AI demo: this profession is finished.

Why the prediction missed

Jobs are bundles, and the automated task is one strand

Reading the scan is the visible task, so it stood in for the job. But the job is also choosing which study to order, protocoling it, catching the incidental finding nobody asked about, talking to the surgeon, doing the biopsy, and holding legal responsibility for the answer. Automating one strand of a bundle doesn't eliminate the bundle — it changes the mix. This is the oldest finding in automation economics and it gets rediscovered, with surprise, every cycle.

Cheaper reads mean more imaging, not fewer radiologists

When a task gets faster and cheaper, demand for it tends to expand — the same dynamic that made more efficient steam engines burn more coal, not less. Medicine is close to a textbook case: there is enormous latent demand for imaging, an aging population supplies more of it every year, and any capacity freed by software gets absorbed by volume. Efficiency gains became throughput, not layoffs.

The bottleneck moved

Clearing a model for one narrow finding is not deploying it across a health system. Liability, malpractice insurance, reimbursement codes, workflow integration, and the question of who gets sued when the model misses — each became the new constraint the moment the technical constraint relaxed. In regulated, liability-heavy industries, the last mile is measured in years per use case, and it doesn't compress just because the model improves.

What actually changed

Radiology didn't ignore AI — it metabolized it. Triage tools flag likely bleeds so urgent scans jump the queue. Detection aids act as a second reader. Measurement and dictation got faster. The realistic near-term shape of the profession is a radiologist reading more studies with algorithmic assistance, and departments buying software the way they buy any productivity tool: cautiously, per use case, with a lawyer in the room.

The reusable checklist

None of this proves AI won't displace work — it plainly will, unevenly, and some of it already has. What radiology gives you is a test to run on any "profession X is finished" claim before you believe it, invest on it, or panic about it. Ask four things: What share of the job is the automatable task, honestly measured? What happens to demand when that task gets cheap? Who carries legal liability for the output, and will they accept a model carrying it? And who actually purchases the software, through what approval process, on what timeline? The confident predictions skip all four. The 2016 one skipped all four, made by one of the smartest people in the field, about the most exposed specialty there was. That should set your prior for the current crop.