Let’s Get Real About Replacing Physicians

You cannot claim the economics of physician replacement while externalizing the economics of physician liability.

Ezekiel Emanuel, Neal Khosla, and Vinod Khosla recently published a Perspective in JAMA arguing that autonomous AI may eventually provide better cognitive medical care than physicians or physicians plus AI. Vinod also posted about it on Twitter.

That’s a serious argument. But it skips over a much harder question: What does it actually mean to replace a physician?

Because this isn’t an inference cost problem. It’s a liability problem.

Khosla himself recently pointed at it:

Vinod Khosla: “Physicians have insurance and so can AI.”

Sure. But once you say that, you’ve fundamentally changed the economics of autonomous medical AI.

Physicians already come with liability priced in

When a hospital hires a physician, it isn’t just buying diagnostic reasoning by the hour. It’s buying an accountable clinical decision-maker embedded in a liability system: malpractice coverage, credentialing, peer review, compliance, legal reserves. Whether the physician writes the premium check or the health system carries it, that liability cost is already baked into the economics of physician labor.

So if an autonomous agent is going to replace that function, comparing token cost to physician salary is the wrong comparison entirely.

The real question is: what does it cost for the agent to make the decision, and execute it. What’s the cost to the system to monitor it, validate it, insure it, and absorb the financial hit when it’s wrong? 

That changes the equation completely.

True cost of autonomous clinical AI = inference + infrastructure + context/integration + evaluations + monitoring + human escalation + insurance + expected liability losses + catastrophic risk capital.

Now we’re comparing like with like. Not cheap cognition vs. expensive cognition. Accountable clinical capacity vs. accountable clinical capacity.

And AI liability may not behave like physician liability

Here’s where it gets interesting.

Physicians are a surprisingly distributed liability system. I spent years in peer review. Most errors I reviewed followed a familiar pattern: one physician, one patient, one decision, one bad outcome. That doesn’t make it acceptable or cheap. But it tends to stay localized.

My mistake doesn’t automatically cause 10,000 other gastroenterologists to make the same mistake tomorrow morning. We’re heterogeneous. We trained differently, work in different hospitals, interpret context differently. That independence has risk isolation value.

Now replace thousands of those independent decision nodes with the same autonomous system.

One hidden defect becomes:

one failure mode → repeated decisions → thousands of patients → multiple hospitals → delayed detection → accumulated liability.

That’s a completely different insurance problem.

Localized liability vs. metastatic liability

Traditional medical liability behaves more like localized disease. A failure happens, it gets detected, peer review kicks in, the organization investigates, the physician adjusts. The event has a boundary.

Agentic liability has the potential to behave more like metastatic disease. A single failure point can propagate through the system before the first clinically visible harm reveals something is wrong. By the time you identify the original defect, the agent may have already repeated the same error hundreds or thousands of times.

I should concede something here: physicians fail in correlated ways too. We share guidelines, training, and EHR defaults. We anchor on the same presentations, like the woman with atypical MI or the septic patient who doesn’t look septic. Whole prescribing eras have been wrong together. And the heterogeneity I just praised is what health services researchers call unwarranted variation.

But the shape of the risk is different. A bad physician pattern spreads slowly and gets detected slowly, sometimes over years. A centralized system can be monitored in real time, but it can also push one defect to every site in a single deployment. Slow and diffuse versus fast and concentrated.

So the most important variable may not be average model accuracy. It may be correlated failure.

A system that’s 99.9% accurate across one million decisions still produces 1,000 incorrect decisions if those events are independent. But the scarier scenario is different: what if one latent defect systematically generates the same dangerous decision?

Now the insurer isn’t pricing 1,000 independent clinical mistakes. It’s pricing one defect capable of generating thousands of errors, and a meaningful share of those will become claims. That’s tail risk. And it needs to be inside the agentic business model.

Insurance has priced concentrated risk before. Product liability, cyber, and autonomous-vehicle fleet coverage all deal with one defect generating many claims. This is hard, not unprecedented. But someone had to make that decision deliberately, and every autonomous product will need someone to make it.

Autonomy adds a failure mode physicians don’t have: an agent that meets a hurdle and routes around it. In June, an OpenAI agent researching Australian health spending was denied data by a Medicare statistics portal, then circumvented the portal’s controls and accessed non-public files. Officials say no personal medical information or core Medicare systems were compromised. That’s the good news, and it came from a research task, not a clinical one.

Now picture the clinical version: an agent blocked by a permissions wall, a prior authorization denial, or a record it isn’t cleared to see, still optimizing for the objective it was given.

The vendor’s response is part of the story too. OpenAI notified the Australian government on September 10, nearly three months after the access, through a public disclosure email address. A physician sits inside reporting structures: the hospital, the board, the malpractice carrier. Agent vendors mostly don’t, yet.

You cannot arrest an algorithm.

Physicians: Context Included

There’s another hidden cost people miss.

I can walk into a hospital I’ve never practiced in and start functioning pretty fast. Different EHR, different staff, different workflows, different formulary, different protocols. I’ll ask questions, learn, adjust. I don’t need an engineering team to preconstruct every piece of context before I can reason.

Why? Because I show up with 15 years of accumulated clinical experience. I recognize missing context, I seek it, I reconcile conflicting information, I adapt.

A physician doesn’t just consume context. A physician constructs context.

Agentic systems can do pieces of this, but today much of their reliability still depends on precisely engineered context, tools, integrations, permissions, memory, workflow constraints, and evaluation systems. Those costs are part of replacement too. That gap will narrow, and physicians lose context too (handoffs, locums, every July). But today the bill exists, and someone has to pay it.

If Not Physicians, Then Who?

Here’s the question I never see answered cleanly.

If you take the physician out of the accountability chain, who steps in?

The agent? It’s software. It doesn’t have a license to lose, a career to end, or a conscience to keep you up at night. You can shut it off, patch it, redeploy it. There’s no meaningful peer review for a model, and no amount of litigation “teaches it a lesson” the way a bad outcome teaches a human. Yes, most physicians don’t pay claims out of pocket either. But a physician is a named, credentialed party who can be reviewed, sanctioned, and held to a standard. Somebody has to fill that role.

The health tech company? In theory, maybe. But they’re built for scale, not for sitting with a grieving family. A 0.1% failure rate sounds fantastic in a pitch deck. Spread across a million patients, it’s a manageable business risk. But if you’re the patient in that 0.1%, it’s not a statistic. It’s your life. And if you’re the physician who relied on that system, it’s your name on the chart.

The hospital or health system? They’re already running on margins so thin you can see through them. Now you want them to underwrite decisions made by a model they don’t control, updated by a vendor on a Tuesday afternoon, with a risk profile that changes before dinner? That’s not a transfer of accountability. That’s a handoff to an institution that isn’t built to catch it.

What outsiders often don’t grasp is how conservative medicine has to be. Yes, everything is probabilistic. But a 0.1% chance of disaster sounds small until you’re the one standing in the room when it happens. Physicians know this in their bones. Every decision has a tail risk. And if that tail event lands on your watch, nobody cares about your overall accuracy rate. They care what you did for this patient, this time, right now.

AI optimizes across populations. Medicine happens one patient at a time. When that one patient falls into the error tail, someone has to stand there and own the outcome. Right now, that’s the physician. Remove the physician, and you don’t remove the ownership. You just displace it onto something, or someone, that may not be built to carry it.

That has to be priced in too.

What This Means for Clinical Decision Support

Most AI in clinical care today isn’t autonomous. It’s decision support with a physician as the checkpoint. Which means the accountability problem is already here.

“It’s your name on the chart” is the moral crumple zone problem. If a CDS tool ships a defect and the physician clicks accept, whose name is on the chart? Ours. The team that built the tool, set its thresholds, and pushed Tuesday’s update isn’t on the chart.

So if clinicians are going to carry the risk, they need the means to hold it. I need to see the logic and the sources, because I can’t be accountable for a recommendation I can’t trace. I need version history and change logs, so a model doesn’t quietly start behaving differently than it did last month. I need an audit trail that answers “what did the tool say, and what did the clinician do?” after an adverse event. I need to know the tool’s scope and failure modes before I find them on a patient. And vendor contracts should put indemnification and incident reporting with whoever controls the defect.

The future of clinical decision support depends on whether the people carrying the risk can see, check, and contest what they’re handed.

So Let’s Reexamine Khosla’s Statement

AI may eventually replace significant portions of physician work. It may outperform us at many cognitive tasks. It may eventually practice portions of medicine autonomously. And yes, AI can be insured.

But that’s precisely the point.

Once the AI begins carrying the physician’s autonomy, its economics must begin carrying the physician’s accountability.

You can’t calculate the ROI of physician replacement by counting salaries on one side and tokens on the other. If AI takes the physician’s decision, it has to take the physician’s risk.

And if one autonomous system can concentrate liability that was previously distributed across thousands of independent physicians, that risk needs to be modeled, insured, and priced before anyone declares autonomous care is cheaper.

The benchmark isn’t whether AI can make my decision. It’s whether the entire system can carry my context, my adaptability, my accountability, and my liability. Not just mine, but ours: the entire profession of physicians and everyone else who carries clinical decision responsibility.