To truly grasp how artificial intelligence will impact the world, look beyond coding, law, or finance, and focus on healthcare. This sector presents AI with its most formidable challenge: navigating layers of regulation, managing life-or-death stakes, interpreting complex biology, and addressing a deeply human, compassionate core that many assume is beyond the capacity of a machine to replicate.

Nearly a decade ago, computer scientist and Nobel Prize-winner Geoffrey Hinton, known as the “Godfather of AI,” suggested that hospitals cease training radiologists, predicting that AI would surpass human capability within five years. Almost ten years later, the radiology profession is larger than ever. Among the artificial intelligence and machine learning tools granted FDA approval between 1995 and 2024, 723 were radiology devices. While the technology has advanced, the human practitioners have not departed.

When I recently discussed this with Hinton, he was quick to reframe his position rather than retract it. He clarified that his miscalculation was not regarding the technology itself, but rather the economics.

“Healthcare is a highly elastic market,” he explained. “If you enable a healthcare worker to accomplish ten times as much, we will simply consume ten times as much healthcare. Older individuals, in particular, have an inexhaustible capacity for it.”

The prevailing question—“Will AI replace doctors?”—is fundamentally flawed. The demand for healthcare is virtually limitless. There is always another scan to interpret, another condition remaining undiagnosed due to time constraints. AI will not diminish the medical workforce; instead, it will reveal the vast extent of unmet need that has always existed.

Instances where AI surpasses doctors, and where it falls short

In certain environments, AI is already outperforming physicians. Cardiologist and researcher Eric Topol noted scenarios where autonomous AI systems outperformed doctors utilizing AI as an assistive tool. “I still maintain that the combined approach is likely to prevail,” Topol stated. “However, my confidence is not as high as it was in 2019.”

Why might standalone AI occasionally outperform a human aided by AI? One reason is a phenomenon researchers term “automation neglect,” where physicians adhere to their initial diagnosis and fail to adjust, even when the system proposes an alternative. Another possibility is that we have not yet mastered effective collaboration with these tools.

Not all evidence supports the machine’s superiority. In a randomized controlled trial published in a medical journal, a cardiologist and colleagues evaluated an AI system on complex cardiology cases involving a suspected condition that challenges even experienced clinicians.

“Specialists are scarce,” he remarked. “Could AI assist generalists in thinking like specialists?”

It could. General cardiologists supported by AI generated assessments that specialist reviewers favored, featuring fewer clinically significant errors. However, 6.5% of the AI’s responses included clinically significant hallucinations.

The subsequent events made this finding particularly valuable. “When the human cardiologist challenged the AI model, asking, ‘Are you certain the echocardiogram showed a thickened ventricle?’ the AI would correct its error.” The machine was unaware of its mistake until prompted by a human.

There are also warning signs. Just last month, Topol observed that a paper in a recent journal evaluated medical triage using ChatGPT’s most advanced model. It misdiagnosed the urgency more than half the time, instructing patients who required emergency care to stay home. “We still have a long way to go,” he commented.

The evidence is inconsistent. For certain tasks, AI on its own is superior. For others, the collaboration of human and machine outperforms both individually. In other cases, the technology proves dangerously unreliable. The true challenge is not whether AI functions, but knowing when to use it.

Transitioning from reactive to preventive care

The most profound change may lie not in diagnostic precision, but in timing. Contemporary health systems are designed to treat illness after symptoms manifest. Topol believes AI could facilitate a shift toward upstream medicine.

“The three primary age-related diseases—neurodegeneration, cancer, and cardiovascular disease—all incubate within our bodies for 15 to 20 years,” he explained. “We have a substantial window of opportunity to address them, but previously lacked the means to integrate all the data. We didn’t even possess all the data.”

Now, we are beginning to do just that. Wearables and other devices generate continuous streams of data regarding heart-rate variability, blood oxygen, and sleep. Researchers at Stanford recently demonstrated that 130 conditions could be accurately predicted using this data. Various tools can now estimate key health metrics. According to Topol, the missing component is the immunome, a comprehensive map of an individual’s immune function.

“Following the brain, the immune system is the body’s most complex system,” he stated. “And we currently have no clinical method to measure it. In 2026, that is unacceptable.”

He posits that a deregulated immune system is the common link between cancer, neurodegeneration, and heart disease, and that measuring it will usher in a new era of risk prediction.

The opportunity lies not in substituting doctors with a single revolutionary product, but in constructing the infrastructure for a new upstream model of preventative care involving sleep, wearables, and blood proteins. The true potential of AI may be its ability to silently monitor the body’s earliest warning signals and intervene well before illnesses become apparent.

The legal, ethical, and human boundaries of AI in healthcare

The integration of AI in healthcare, however, will not be solely a technical endeavor. Hinton highlighted a legal imbalance. If a physician neglects to use an available AI tool and a patient dies, no lawsuit is filed. Conversely, if a doctor uses AI and harm ensues, liability may be immediate. This system discourages early adoption.

Meanwhile, human error remains widespread. “We are aware of at least hundreds of thousands of cases in the U.S. that result in numerous deaths,” Topol informed me. “Yet, we rarely discuss this. We continue to focus on the mistakes made by AI.”

Additionally, the issue of empathy remains unresolved. When I asked Hinton if he would be comfortable receiving end-of-life care from AI, he hesitated. “I might suspect it was feigning empathy,” he replied. He then added: “But I believe AIs can possess genuine empathy.”

Topol holds a different view. “AI is excellent at simulating empathy,” he stated. “However, a machine cannot truly comprehend empathy. Patients desire to look someone in the eye and feel that person cares for them. That is the essence of medicine. No machine will ever be able to fully replace that.”