If autonomous AI wins, the hard part isn't the medicine

Estimated reading time: 3 minutes

New here? Subscribe to AI Dx, one issue every Saturday: https://bhargavpatelmd.beehiiv.com

TL;DR

A JAMA Perspective argues autonomous AI may soon deliver better care than physicians, and than physicians using AI.

Even if you grant that, the capability question isn't the hard one. Four others are.

Liability: remove the licensed human who signs off, and malpractice has nothing to attach to. Reimbursement: payment codes are built around a clinician's work. Care no human delivered has no slot. Regulation and education: the FDA leans on clinician review; training assumes skills AI may erode.

Capability is racing. The infrastructure around it is barely moving, and that gap is the risk.

There's a provocative claim from a JAMA Perspective: autonomous AI will soon out-practice not just doctors but doctors using AI. The debate has fixated on whether that's true.

I want to do something different. Grant it. Assume the capability arrives. Because the moment you do, you realize the medicine was never the hard part. Everything built around the medicine is, and almost no one is working on it.

The scoreboard was never the real question

"Can AI out-diagnose a doctor" is a benchmark question, and benchmarks are the part of this we're good at measuring.

The questions that decide whether patients benefit (or get harmed) sit downstream, in the legal, financial, and institutional scaffolding medicine runs on. That scaffolding was built for human clinicians. Here's what happens when you remove them.

1. Liability has nothing to attach to

Today, when an AI-assisted decision goes wrong, the answer to "who's responsible" is the physician who signed off. The whole malpractice framework rests on a licensed human's judgment. Remove the human and it has no anchor.

You can't sue an algorithm. Is it the vendor? The hospital? The clinician "supervising" thousands of autonomous decisions? No one has solved this, and it's the mechanism by

which harmed patients get made whole.

2. Reimbursement has no slot for it

We pay through codes tied to a clinician's work. If an autonomous system manages a disease better and cheaper with no clinician in the loop, what gets billed, and to whom? Payment shapes behavior.

Until there's a way to pay for autonomous care that rewards outcomes over volume, the economics will quietly steer whether these tools get used at all.

3. Regulation wasn't built for a non-human practitioner

The FDA clears many decision-support tools partly because a clinician reviews the output.

That reviewer is load-bearing. Take them away and you're regulating an autonomous practitioner that isn't a person, can't hold a license, and updates continuously.

The recent FDA discussion paper on generative AI is an early attempt, but it's a starting question, not an answer.

4. Education is training for a job that's changing underneath it

If AI handles more of the cognitive load, what do we train physicians to do, and how do they stay sharp at skills they use less?

Deskilling isn't inevitable — aviation designed around it — but it isn't automatic to prevent either. Right now medical education is largely preparing physicians for a version of the job that's shifting faster than the curriculum.

The gap is the risk

The capability question is advancing fast and will keep advancing. The infrastructure questions (liability, payment, regulation, training) are barely moving.

That gap is where the danger lives. Not in whether the AI is good enough, but in dropping capable autonomous systems into a structure built for humans and never rebuilt.

Capable technology in unprepared infrastructure doesn't produce safe care. It produces confusion about who's responsible when something goes wrong.

Final thought

The future of autonomous medical AI won't be decided by a benchmark score. It'll be decided by whether we do the unglamorous work (rebuilding liability, payment, regulation, and training) before the capability fully arrives, not after a patient is harmed and no one can say who's accountable.

The scoreboard is the easy part. The harder question: which breaks first if autonomous AI pulls ahead — liability, reimbursement, regulation, or training?

Is your organization preparing for AI that delivers care, or still thinking about AI that advises on it?

—

Bhargav Patel, MD, MBA

Physician-scientist, researcher, and leader at the intersection of clinical medicine and technology. Former Chief Medical Officer and advisor at top Health AI companies. Founder of AI Dx & Sage Psychiatric Professionals.

Subscribe to AI Dx: https://bhargavpatelmd.beehiiv.com