If you want to understand how artificial intelligence actually impacts the world, look no further than coding, law, finance, and more. Look at healthcare. That’s where AI will face its toughest challenges. Multiple layers of regulation, life-and-death stakes, complex biology, and a deeply human and caring core that most people think cannot be replicated by a machine.
Nearly a decade ago, computer scientist and Nobel Prize winner Jeffrey Hinton (known as the “Godfather of AI”) said hospitals should stop training radiologists because AI would do a better job within five years. Almost a decade later, there are more radiologists than ever before. Of the 950 artificial intelligence and machine learning tools that received FDA approval between 1995 and 2024, 723 were radiology devices. The machine has been improved. The humans didn’t leave.
When I raised this issue with Hinton recently, instead of backing down, he immediately reconsidered. It wasn’t the technology that misjudged him, he says. It was economics.
“The health care market is a very resilient market,” he told me. “If we allow health care workers to provide 10 times more care, we all receive 10 times more care, especially the elderly, who can absorb unlimited amounts of care.”
The standard question “Will AI replace doctors?” turns out to be wrong. The demand for healthcare is virtually limitless. Another scan always needs to be read and another condition goes undiagnosed because no one has time to look into it. AI will not shrink the healthcare workforce. It becomes clear that unmet needs have always existed.
When AI will outperform doctors and when it will fail.
In some situations, AI has already surpassed doctors. Cardiologist and researcher Eric Topol pointed to five studies in which AI systems working independently outperformed doctors who used AI as a tool. “I still think this combination has a good chance of winning,” Topol told me. “But I’m not as confident as I was in 2019.”
Why does AI on its own sometimes perform better than humans with AI assistance? One explanation is what researchers call automation neglect. Doctors stick to their initial diagnosis and fail to adjust even when the system suggests alternatives. The other is that we simply haven’t learned how to use these tools effectively.
Not all the evidence is in favor of the machines. In a randomized controlled trial published in natural medicineCardiologist Jack Wosullivan and his colleagues tested the AI system on complex heart cases suspected of genetic cardiomyopathy, which is difficult to diagnose even for experienced clinicians.
“There’s a shortage of experts,” he says. “Can AI help generalists think like them?”
It’s done. AI-assisted general cardiologists produced ratings with fewer clinically significant errors and preferred by expert reviewers. However, 6.5% of AI responses included clinically significant hallucinations.
What made this discovery useful is what happened next. “When a human cardiologist asked the AI model, “Are you sure that the echocardiogram showed ventricular thickening?” the AI would automatically correct it.” The machine didn’t realize it was wrong until someone asked.
And there are some signs to look out for. Just last month, Topol said: natural medicine We evaluated medical triage using ChatGPT’s state-of-the-art model. More than half of the time, patients were triaged incorrectly, with patients in urgent need of an emergency room being told to stay home. “We have a long way to go,” he said.
The evidence is not uniform. For some tasks, AI alone performs best. For others, humans and machines together outperform one or the other. Additionally, this technology can be dangerously unreliable. The real issue is not whether AI works. It’s about knowing when.
Transition from reactive to preventive care
The most important change may not be in diagnostic accuracy but in timing. Modern medical systems are built to treat diseases after symptoms appear. Topol believes AI could help move medicines upstream.
“The three major age-related diseases, neurodegeneration, cancer, and cardiovascular disease, all have an incubation period of 15 to 20 years in our bodies,” he told me. “We had such a great runway available, but there was no way to integrate all the data. We didn’t even have all the data.”
Now we are starting to do so. Half a billion people already use smart watches and other wearables, which generate a continuous stream of heart rate variability, blood oxygen, and sleep data. Researchers at Stanford University recently showed that they can accurately predict 130 conditions from a single night of sleep sensor data. Organ clocks derived from thousands of blood proteins now allow us to estimate the biological age of individual organ systems. The missing piece, Topol says, is the immunome, a comprehensive map of a person’s immune system.
“Next to the brain, the immune system is the most complex system in the body,” he said. “And there’s no way to measure that in the clinic. In 2026, that’s scary.”
He believes that dysregulation of the immune system is a common thread connecting cancer, neurodegeneration and heart disease, and that measuring it could open up a new era of risk prediction.
The opportunity lies not in replacing doctors with a single breakthrough product, but in building an infrastructure around new upstream models of preventive care, such as sleep, wearables, and blood proteins. The real promise for AI may be to silently monitor the body’s early warning signs and intervene long before disease becomes visible.
Legal, ethical, and human limits of AI in healthcare
However, the implementation of AI in healthcare is not purely technical. Mr. Hinton pointed out the legal asymmetry. If a doctor fails to use available AI tools and a patient dies, no one will be sued. However, if a doctor uses AI and harm occurs, he or she could be immediately held liable. This system discourages early adoption.
On the other hand, human error remains prevalent. “We know that there are at least 12 million diagnostic errors in the United States each year, resulting in approximately 800,000 people being disabled or dying,” Topol told me. “And we don’t tend to talk about it. We keep talking about the mistakes that AI makes.”
Furthermore, the issue of empathy remains unresolved. When I asked Hinton if he would feel safe being cared for by an AI at the end of his life, he paused. “You might think it’s fake,” he said. He added: “But I think AI can be really empathetic.”
Mr. Topol disagrees. “AI is very good at conveying empathy,” he told me. “But no machine will ever know what empathy is. People want to look someone in the eye and know that that person cares about them. That’s what medicine is about. No machine will ever truly replace that.”
