AI Is Changing Radiology Jobs — But Not How Hinton Predicted

In 2016, Geoffrey Hinton — the Nobel Prize-winning "godfather of AI" — declared that training new radiologists was pointless. Within five years, he predicted, computers would read scans better than any human. Last year, he told the New York Times he was wrong.

But here's what the correction doesn't tell you: being wrong about replacement isn't the same as nothing changing. A December 2025 Stanford preprint by Curtis Langlotz and colleagues — the most methodologically explicit workforce analysis to date — projects a 33% reduction in radiologist hours over the next five years, with the range running from 14% to 49%. Langlotz also concluded that radiologist job loss is unlikely in the foreseeable future, because imaging demand keeps growing faster than the workforce can absorb it.

The threat is real. It's just not what was advertised. Hours per study are being compressed; the profession isn't in immediate freefall. What determines which side of that gap you land on is specific, learnable, and illustrated by two radiologists who already made their choice.

Two Radiologists, One AI Wave, Very Different Bets

Jeff Chang entered medical school at New York University at age 16 and became the youngest radiologist in U.S. history. He spent the next decade working overnight emergency shifts, and the toll was concrete. "When I do two or three weeks of night shifts in a row," he said, "it's just like your brain turns to mush." He didn't lobby for shorter shifts. He went back to school for machine learning, identified that 75% of a radiologist's time was spent dictating reports, and co-founded Rad AI in 2018. The company now works with more than 40% of U.S. health systems, was valued at $528 million in 2025, and deploys AI that saves radiologists roughly an hour per shift on report generation. Chang still practices emergency radiology.

Nina Kottler took the other path. A practicing radiologist with a graduate degree in applied mathematics, she became the first radiologist to join Radiology Partners in 2013, built its AI program from research pilots through national deployment, and was promoted in late 2025 to Chief Medical AI Officer of Mosaic Clinical Technologies — the technology arm overseeing AI governance across more than 50 million annual exams. She did not write a single line of model code. She governed the humans and systems that used it.

Two paths, both viable: build the tools, or become the clinical expert who deploys and governs them at scale. Neither required abandoning the identity of being a radiologist. Both required identifying, earlier than their peers, which part of the job was about to be restructured.

For anyone who isn't a coder and isn't interested in founding a startup, Kottler's path is the more immediately actionable model. It requires no career change — only a deliberate shift in how expertise gets directed.

The Job Isn't Disappearing. It's Splitting.

To know which bet to make, you need to know which specific tasks are actually being automated — and which ones just became more valuable because a machine is now making the first call.

On the automating side, report drafting is the single largest target. A longitudinal multireader study published in the Journal of the American College of Radiology in early 2026 found that AI-generated preliminary chest-radiograph reports reduced mean reading time from 25.8 to 19.3 seconds per case, with acceptance rates climbing above 60% — and nearly 69% for normal studies. Scan queue prioritization is moving just as fast: a 2026 study across more than 313,000 emergency-department visits found that an AI dispatch policy cut the 90th-percentile wait time for clinically actionable CT studies by 43 minutes. Routine protocoling and follow-up coordination are being automated in practices like the one Chang's company built, which identifies follow-up recommendations and communicates directly with patients, increasing completed follow-up appointments by up to 40% in company-reported data.

The goal is not a robot radiologist. It's a radiologist whose clinical expertise is complemented by AI to enable better patient care, information retrieval, improved productivity, better accuracy, more accurate reporting and more effective communication.
— Tessa Cook, Associate Professor of Radiology, University of Pennsylvania

On the "becoming more valuable" side, the evidence is just as striking — but the dynamic is counterintuitive. When AI gives a wrong answer, radiologists who maintain independent judgment are essential; those who defer to the machine become dangerous. A 2026 Radiology study using eye-tracking technology found that when an AI system missed a cancer on a mammogram, radiologist sensitivity dropped from 71% to 39% — a 32-percentage-point collapse attributable to reduced visual search effort when the machine was silent. The radiologists who caught the machine's misses were the ones who kept looking even when the AI flagged nothing.

In pulmonary embolism detection across more than 32,000 CT angiography scans at a large hospital network, radiologists and AI agreed 97.79% of the time. But when they disagreed, expert adjudicators sided with the radiologist nearly 89% of the time. The human wasn't redundant. The human was the error-correction layer.

Langlotz frames it precisely: "Machine intelligence and human intelligence are different kinds of intelligence." AI can scan every pixel without fatigue; a radiologist understands why a finding matters for this patient, in this clinical context, right now.

Chang built tools for the first column. Kottler built a career around governing the second. The radiologist who is neither building nor governing is the one with the most uncertain position.

This same logic applies across any knowledge-work field where AI is beginning to handle the high-volume, structured output. The question is always the same: do you build the pattern-recognition machine, or do you become the person responsible for what happens when it fails?

Will There Actually Be Fewer Jobs?

Knowing which column you're in is clarifying. But there's a harder question underneath it: if AI is genuinely compressing radiologist hours, does that eventually mean fewer radiologist jobs — and what do the actual labor-market signals say right now?

In early 2026, Mitchell Katz — CEO of NYC Health + Hospitals, the largest public hospital system in the United States — said publicly that his system was ready to replace radiologists with AI "if we are ready to do the regulatory challenge." Radiologists pushed back hard, and the clinical counterarguments are legitimate: current AI cannot match a radiologist's contextual judgment, integration is genuinely difficult, liability remains with the physician, and autonomous AI has not been cleared for general diagnostic use. But the fact that a major hospital CEO said it publicly signals that the institutional pressure is structural, not hypothetical.

Now we not only need to at least maintain or improve quality, but more importantly, we need to improve the efficiency and workflow of physicians, because there are not enough radiologists to do the amount of imaging that is coming in.
— Nina Kottler, Chief Medical AI Officer, Mosaic Clinical Technologies

The clearest substitution evidence comes from breast screening. A large 2026 UK study found that using AI as the second reader in a double-reading workflow reduced total reader time by 32% while increasing cancer detection rates. A separate economic model found that replacing one human reader with AI saved roughly £31 per woman screened. These are modeled results, not implemented layoffs. But they are the math that procurement departments will eventually run.

Against this: a 2026 job-posting analysis of more than 20,000 U.S. radiology listings found that approximately 1,470 active openings had been unfilled for more than two months, with the shortage concentrated in rural and low-population states. The UK's Royal College of Radiologists reported a 30% shortfall of consultant radiologists, projected to reach 39% by 2029. When demand is growing faster than supply, productivity tools are more likely to be used to clear backlogs than to eliminate positions — at least in the near term.

The Langlotz forecast and the breast-screening evidence describe a real compression of human labor per study. Whether that compression translates into fewer jobs depends on whether imaging volume grows fast enough to absorb the productivity gain — and right now, it is. That may not be permanent. The window to reposition is open while the shortage still provides cover, not after it resolves.

In any profession facing AI-driven productivity compression — accounting, legal review, software testing, content moderation — the same dynamic holds. Current labor shortages create a buffer that makes the near-term employment picture better than the task-exposure math would suggest. But that buffer is temporary, and the repositioning window is now.

What to Do This Week

Jeff Chang and Nina Kottler didn't have better information about AI than their peers. They had the same uncertainty everyone else had in 2016, 2018, 2020. What they had was a willingness to act on it before the answer was obvious.

Hinton was wrong about replacement. He may still be right that the job will never be the same. The radiologists who are positioned well for what's coming are not the ones who were confident it would be fine — they're the ones who were worried enough to look carefully at which parts of their work a machine was already doing adequately, and then deliberately moved their attention somewhere else.

Start with one task audit this week: pull up your last ten cases and sort them by the column they belong in. How many were routine, normal, high-volume reads where AI acceptance is already above 60%? How many required the kind of contextual judgment, clinical communication, or disagreement-resolution that the evidence shows machines still get wrong? That ratio is your personal exposure map — and it's more useful than any forecast.

The radiologists who are anxious right now are paying attention. That's not the problem. That's the starting point.

Recommended Tools & Resources