We may someday look back on this moment the way we now look back on paper medical charts.

Primitive.

A study recently published in Science found that an OpenAI reasoning model outperformed experienced emergency physicians when diagnosing patients using information available in their electronic medical records. Not on textbook cases. On real patients treated in busy emergency departments.

The researchers at Harvard Medical School and Beth Israel Deaconess Medical Center intentionally tested the AI under the same imperfect conditions physicians face every day: incomplete histories, evolving symptoms, and electronic records that rarely tell the whole story.

In one case, the AI correctly connected a patient’s worsening condition to a previously documented history of lupus—a diagnosis the treating physicians missed.

Why This Result Isn’t Actually Surprising

That result shouldn’t surprise us as much as it once might have.

Medicine has become extraordinarily complicated. No physician—not even an excellent one—can remember every disease, every study, every diagnostic pathway, or every obscure connection buried somewhere in a patient’s chart. We shouldn’t expect them to. Emergency physicians make hundreds of consequential decisions under enormous time pressure, often caring for patients they’ve never met before.

The question, then, isn’t whether AI is smarter than doctors.

It isn’t.

The better question is whether ER doctors can use AI as a tool to be even better.

That has happened before. CT scanners didn’t replace physicians. Neither did MRI machines, electronic medical records, computerized drug interaction alerts, or countless other advances. They became tools that helped physicians make safer decisions.

AI may simply be the next tool.

What AI Can’t Do—And What It Never Forgets

The model studied in Science certainly has limitations. It couldn’t examine a patient, interpret body language, recognize subtle physical findings, or earn a patient’s trust. Those are uniquely human skills, and they remain central to good medical care.

But it also doesn’t become fatigued twelve hours into a shift. It doesn’t anchor on its first diagnosis. It doesn’t overlook a note written three years ago because the emergency department is overflowing with patients.

If technology can reliably help physicians avoid preventable diagnostic errors, hospitals should be asking an important question today.

When does that technology become part of reasonably safe medical care?

That’s ultimately how medicine advances. We learn. We develop better tools. We discover better ways to care for patients. And over time, yesterday’s innovation becomes tomorrow’s expectation.

What Patients Deserve as This Technology Arrives

Missed diagnoses remain one of the leading causes of serious patient harm. Sometimes the cause is individual error. More often, it reflects the reality that modern medicine has become too complex for any one person to navigate perfectly every time.

If AI can help physicians deliver safer care—and the evidence increasingly suggests it can—patients deserve healthcare systems that thoughtfully evaluate, appropriately implement, and carefully monitor those tools.

Technology changes.

Our expectation that patients receive reasonably safe medical care should not.