Who Takes the Blame When AI Gets It Wrong in Medicine?

I’m a doctor, and I’ve been involved with AI for decades. I’ve watched it evolve from CNNs to reinforcement learning and beyond.

I’ve always said there’s no way AI can completely replace human doctors.

The idea that “AI alone is better than AI + doctor” is false.

A lot of people point out that there are bad doctors, doctors who make big mistakes, doctors who don’t know their patients, and that society is losing faith in the profession.

I know there are bad, corrupted, and dishonest doctors out there.

Now imagine the same problems with AI: bad AI, corrupted AI, dishonest AI. Who takes accountability?

Hypothetically, everything is done by AI, including the robotic surgeon.

You walk into the ER with chest pain. You’re triaged by AI, blood is drawn by AI, and after three hours the AI discharges you based on every PubMed paper and universal EHR data. You disagree with the decision because you still have chest pain, but AI reassures you based on complex computed data (that also assumes generalization). A few weeks later, you have a massive heart attack and permanent heart muscle damage.

Who takes the blame? The AI? How?

Liability remains the biggest challenge in the AI economy.

I’ve been saying we need AI insurance and AI audits.

For the next 5 to 10 years, I see AI handling telehealth and basic primary care, including annual visits and straightforward cases.

But when it comes to patients with complicated comorbidities, AI still cannot match a human doctor.

AI systems are extraordinarily good at recognizing patterns across vast datasets and generating statistically likely answers. But patients with complicated comorbidities do not present as clean statistical averages. They present as unique, messy, incomplete, and often contradictory combinations of disease interactions, social context, unspoken priorities, and residual uncertainty.

Current AI still operates primarily through statistical correlation and optimization against training distributions. It has no lived accountability, no capacity for genuine moral judgment, and limited ability to reason abductively under sparse or novel conditions the way an experienced physician does. When the data are incomplete, the case is atypical, or the “right” answer depends on values as much as physiology, the human doctor’s contextual judgment and responsibility remain decisive.

That gap is narrowing, but it has not closed.

Ironically, the strongest future isn’t AI vs. doctors.

It may be that physicians who use AI effectively will replace physicians who don’t.

Figure from my previous publication: Krittanawong C, Zhang H, Wang Z, et al. “Artificial Intelligence in Precision Cardiovascular Medicine.” JACC. 2017;69(21):2657–2664.