Radiologists Aren't Disappearing Because of AI, Despite Hinton's Decade-Old Prediction

In 2016, Geoffrey Hinton, one of the pioneers of deep learning, said it bluntly: people should stop training radiologists, because within five years artificial intelligence would be better at the job than they are. A decade later, the number of radiologists in the United States hasn't fallen, it has grown, and new analyses show that AI has entered doctors' offices as a tool they use, not a replacement for them.

A Prediction That Didn't Hold Up

Hinton's 2016 quote has resurfaced in numerous American media outlets in recent months, because ten years on, market data tells a completely different story. Radiologists in the US now earn on average more than half a million dollars a year, and demand for specialists in the field is growing faster than supply. Hinton himself later admitted he had spoken too broadly and hadn't specified that he meant only image analysis, not the entire profession, which also includes talking with patients, planning treatment, and interpreting results in the context of a patient's medical history.

That context turned out to be crucial. Algorithms today can analyze every pixel of an image without fatigue and compare it against millions of prior cases in a fraction of a second, something no human could replicate. But recognizing a pattern in an image is only the beginning of diagnostic work, not the end of it.

The Scale of Adoption in US Hospitals

The US Food and Drug Administration has maintained a list of AI systems cleared for medical use for years, and its latest data show radiology dominates the approvals by a wide margin. More than 1,160 of roughly 1,500 total FDA-cleared AI-based medical devices concern diagnostic imaging, or roughly three-quarters of the entire market. The agency now clears about 30 new algorithms a month, a pace that has been steadily climbing for several years.

The biggest suppliers of these systems are medical equipment giants: GE HealthCare has more than 120 clearances in radiology alone, Siemens Healthineers close to 90, and Philips 50. Alongside them, a growing group of specialized AI startups, like Aidoc and DeepHealth, are building systems focused exclusively on detecting specific pathologies, such as pulmonary embolisms or intracranial hemorrhages.

Artificial intelligence is not a better kind of intelligence, it's a different kind of intelligence. Human plus machine works better than either alone. - Dr. Curtis Langlotz, director of the Center for Artificial Intelligence in Medicine and Imaging, Stanford University

Why Diagnostic Errors Still Happen

The scale of the problem AI is meant to help solve is enormous. An estimated 40 million diagnostic errors occur worldwide each year, and traditional human interpretation of medical images carries a margin of error of roughly 3-5 percent. With millions of X-rays, CT scans, and MRIs performed every year, even a small error rate translates into hundreds of thousands of missed or misread cases.

AI systems can catch some of these mistakes, especially where fatigue, time pressure, or routine make oversights more likely. They don't eliminate the problem entirely, though, and in some cases introduce a new kind of risk: so-called black boxes, neural networks that can't explain why they flagged a particular result. This raises concerns about automation bias, when a doctor trusts an algorithm's suggestion too uncritically, and about complacency, when diagnostic vigilance drops because AI hasn't been wrong before.

A Training Gap

The weakest link in the whole system today turns out not to be the algorithms themselves, but how prepared doctors are to work with them. A 2026 American Medical Association survey found that more than a quarter of US physicians have had no formal training in using AI tools, and just 11 percent have completed a full, comprehensive course. At the same time, a clear majority of respondents say they want more education on the subject, preferably in the form of short modules woven into everyday work rather than separate training sessions run by tech companies themselves.

This gap between how fast AI systems are being deployed and how fast medical staff are being trained isn't unique to the United States. Similar signals are coming from European hospitals, where regulators and medical societies increasingly point out that an FDA clearance or CE mark for an AI device isn't enough if the person using it doesn't understand its limitations.

What This Means for Poland's Healthcare System

In Poland, AI systems for analyzing diagnostic images are already reaching more hospitals and clinics, from oncology to cardiology, and national regulators are only beginning to build oversight frameworks for their use. The American experience points to the direction the European market will likely follow too: not replacing specialists, but equipping them with a tool that speeds up the first stage of analysis while keeping a human as the final decision-maker. The key question is whether Polish hospitals and medical schools can build systematic staff training fast enough, before the number of deployed AI systems outpaces their ability to oversee these tools.