The Future of Radiology Is AI Assisted, Not AI Replaced

In 2016, Geoffrey Hinton, one of the so-called "godfathers of AI" and a Turing Award winner, told an audience something bold: stop training radiologists. "It's just completely obvious that within five years deep learning is going to do better than radiologists," he said. Computers, he predicted, would simply take over the job.

Ten years later, the numbers tell a different story. The United States isn't drowning in unemployed radiologists. It's short of them. Workforce research from the Harvey L. Neiman Health Policy Institute found roughly 37,482 radiologists enrolled to treat Medicare beneficiaries as of 2023, and the gap between how much imaging patients need and how many radiologists exist to read it is expected to persist for decades. Demand for imaging is climbing about 3 to 4 percent a year. The supply of radiologists is growing only about 1 percent a year. Hinton has since acknowledged he got the prediction wrong.

So what happened to the AI that was supposed to make radiologists obsolete?

It turns out AI is genuinely good at reading scans, just not the way Hinton imagined. Across dozens of studies, researchers have found that AI helps the most when it works alongside a radiologist, catching something a tired eye might miss or flagging a scan that needs a second look, rather than replacing that eye entirely. Reading a scan isn't just pattern matching on an image. It's connecting that pattern to a patient's history, symptoms, and the judgment call of what happens next.

That distinction reaches beyond radiology. AI is remarkably good at narrow, well-defined tasks: spotting a suspicious shadow, flagging an anomaly, handling the repetitive first pass. But the moment a decision needs context and consequence, a human still has to be in the room. A decade after the prediction that radiologists were finished, hospitals still need them. They just need them working alongside a new kind of tool, not replaced by it.