In 2016, Geoffrey Hinton — the Nobel Prize-winning computer scientist often called the “Godfather of AI” — told a conference audience that “people should stop training radiologists now.” Deep learning, he said, would surpass human radiologists within five to ten years. He compared the profession to a cartoon character running off a cliff, not yet aware there was no ground beneath them.
I remember hearing about that prediction. I’d already been supporting a radiology practice for several years at that point. My first reaction was uncertainty. Not about whether AI could do it — but about what it would mean for medical imaging groups if it did.
A decade later, CNN published an article calling radiology “the ultimate case study for why AI won’t replace human workers.” The data backs that up. But the real story is more interesting than “AI failed to kill radiology.” AI is transforming radiology. It’s just doing it in a way that makes radiologists more valuable, not less.
The Numbers Don’t Lie
If AI were replacing radiologists, you’d expect falling demand, declining compensation, and empty residency slots. The opposite is happening across every metric:
- The U.S. has roughly 41,000 radiologists, and it’s not enough. The Harvey L. Neiman Health Policy Institute projects the shortage will persist through 2055.
- Half of radiologist job searches in 2023 went unfilled, according to the 2024 AAPPR Benchmarking Report.
- Compensation remains strong. Medscape’s 2025 Physician Compensation Report ranked radiology as the third highest-paid medical specialty — not what you’d expect from a field supposedly being automated away.
- In the 2025 residency match, 97.4% of diagnostic radiology positions and 100% of interventional radiology positions filled. Students aren’t fleeing the field — they’re competing to get in.
- Imaging volumes have grown 31% over seven years, compounding at 4.6% annually.
The Bureau of Labor Statistics projects 5% employment growth for radiology from 2024 to 2034 — above the national average. Meanwhile, the Mayo Clinic — which has deployed over 250 AI models in radiology — grew its radiology staff from about 260 to over 400 since 2016. A 55% increase, during the exact period AI was supposed to eliminate them.
What AI Actually Does in Radiology
From the IT side, I can tell you what AI tools look like when they land in a radiology practice: they’re assistants, not replacements.
The dominant model is what’s called a “second reader.” AI reviews the same images the radiologist reads and flags findings — a potential lung nodule, a suspicious mass, a brain bleed that needs urgent attention. The radiologist makes the final call. Always.
Over 1,300 AI-enabled medical devices now have FDA marketing authorization. Radiology accounts for 77% of all medical AI authorizations. But 97% were cleared through the 510(k) pathway — the FDA’s process for devices that are substantially equivalent to existing tools. That tells you something about how the FDA views them: as instruments, not autonomous diagnosticians.
The most compelling evidence comes from the MASAI trial in Sweden — the first randomized controlled trial of AI in mammography screening. AI-supported screening reduced radiologist workload by 44% while detecting 29% more cancers and achieving 12% fewer interval cancers. That’s not AI replacing the radiologist. That’s AI making the radiologist faster and more accurate simultaneously.
Why Hinton Was Wrong
Hinton himself acknowledged his mistake. In a 2024 New York Times interview, he said he “focused too narrowly on image analysis” and overestimated the pace of AI’s evolution. His revised position: “Most medical image interpretation will be performed by a combination of AI and a radiologist, and it will make radiologists a whole lot more efficient in addition to improving accuracy.”
He was wrong because reading images is only part of what radiologists do. They correlate findings with patient histories. They communicate with referring physicians about ambiguous results. They perform interventional procedures. They make judgment calls that require clinical context far beyond what’s in the pixels.
The image is data. The diagnosis is medicine.
The Honest Caveat
I don’t want to paint this as a simple good-news story, because it isn’t entirely.
A Harvard Medical School, MIT, and Stanford study published in Nature Medicine found that AI’s effect on individual radiologists is highly variable. Poorly performing AI tools actually diminished diagnostic accuracy. And there was no reliable way to predict which radiologists would benefit — experience, specialization, and prior AI exposure didn’t matter.
That study found that in six evaluations of AI in representative clinical environments, 80% showed no change in radiologist performance. The remaining 20% improved. Zero percent worsened overall — but at the individual level, some radiologists performed worse with AI than without it.
There’s also a legitimate concern that productivity gains from AI may benefit health systems and employers more than the radiologists themselves. If AI helps a radiologist read more studies per hour, does that radiologist get paid more — or just expected to do more? Anyone in healthcare knows which way that usually breaks.
And Hinton didn’t say he was wrong about the direction — just the timing. The long-term trajectory of AI capability remains uncertain. What’s true today may not be true in 2040.
What This Means
The radiology story matters beyond radiology because it’s the clearest test case we have for the AI-and-jobs question. This was the profession that was supposed to go first. The most prominent AI researcher in the world said so publicly. And a decade later, the profession is stronger than ever — while AI has become an indispensable tool within it.
The pattern isn’t “AI replaces workers.” It’s “AI amplifies demand.” Imaging volumes are up because AI helps identify findings that lead to more follow-up imaging. AI helps radiologists work faster, which means practices can take on more volume, which means they need more radiologists to review what the AI flags.
From where I sit — in the IT area, not the reading room — AI has created more work for my team, not less. More systems to integrate, more vendor relationships to manage, more infrastructure to support. The same is true for the radiologists I support.
Hinton saw the cliff. He just didn’t see what was being built on the other side.