Microsoft claims its AI is more than twice as accurate as human doctors

Last week, Microsoft published a study in Science. Its AI diagnosed 83% of a set of complex cases from the New England Journal of Medicine. Generalist doctors, with online search and reference material to hand, got 36%.

The headline writes itself: AI beats doctors.

The question I keep coming back to is a different one.

How is a doctor who is right 36% of the time supposed to overrule a system that is right 83% of the time?

On paper, the doctor stays in charge. The AI suggests; the human decides. That's the safety net everyone points to.

But look at what that net asks of the doctor. When the two disagree, the odds say the machine is right. Every time the doctor goes along, they're probably correct to. Every time they push back, they're probably wrong. Within a few months, a sensible doctor stops pushing back.

And then the other 17% has nobody watching it.

That's the real problem. Not that the AI makes mistakes; everything does. It's that a tool this good trains the people around it to stop checking, and checking is the only thing that catches the misses.

There's already a hint of it. In a study of colonoscopies last year, doctors who had got used to AI pointing out growths found fewer of them once it was switched off.

To be fair to Microsoft, the study has limits it states itself. The cases were past records, read as text. The doctors were generalists working alone, not specialists or teams. Nobody has shown yet that it works on a ward.

But better numbers won't make the question go away. They make it sharper.

Being right more often than the machine is no longer the job. Knowing where the machine goes wrong is.

That takes a few unglamorous habits:

• Make your own call before you see its answer, so you have something to disagree with.
• Learn the kinds of cases it gets wrong, not just its average.
• When you overrule it, write down why, and check later who was right.

None of this is only about medicine.

Every business owner now has something that is usually right: a dashboard, a forecast, an adviser, a star salesperson, an AI tool. The better it gets, the less anyone questions it. And the less anyone questions it, the more its rare mistake costs.

The answer isn't to trust it less. It's to keep the muscle that can tell when it's wrong.

What's the "usually right" thing in your business that nobody has questioned in a year?