Why Did They Keep Letting Us Into the House?

August provided two documents that got me thinking. In JAMA, Ezekiel Emanuel and three co-authors argued that autonomous systems will soon beat physicians at five core cognitive tasks. (Then he talked about it again and again.) Days earlier, the AMA and the Digital Medicine Society published a framework naming five enduring physician responsibilities, plus a roadmap asking regulators to protect physician-led care.

Both assume we already know why a patient wants a physician in the room. I don't think we do.

Our professional ancestors showed up at the front door with a lancet, a basin, and enough calomel to loosen a grown man's teeth. People let them in anyway.

Why?

What They Were Actually Buying

Charles Rosenberg argued that heroic medicine worked as explanation. Illness was disequilibrium, and treatment had to produce something the household could see. Purges purged. Emetics vomited. Opium quieted the pain and the bowel. The intervention confirmed the theory in front of everyone in the room, and the physician's standing rested on a shared faith in that theory—which he needed as badly as the patient did.

At the same time, I distrust anyone who gives you one answer to why the door kept opening. A constellation of reasons gets closer: interpretation of bodily suffering. A forecast when certainty wasn't available. Recognition that the illness was real. Permission to take up the sick role, or to put it down. Mediation among the patient, the family, the community, and an approaching death. Ritualized attention to the body. A witness who'd come back. Someone bound by professional norms to stay after treatment failed. Someone willing to make a recommendation and answer for it.

The Hippocratic Prognostic names one of these and treats it as the foundation of the whole practice: cultivate prognosis, because by foretelling at the bedside you'll be more readily believed, so that people will have confidence to entrust themselves to you. becuase people naturally grasp that accurate prognosis requires an understanding of what's going on in the body now and what has happened to the bearer of that body in the days and weeks and years before. That's a business model, written 24 centuries ago, and it places the value on the forecast.

Before I wander off into elder-physician nostalgia, two qualifications.

None of this was ours alone, and much of it was done better by midwives, neighbors, clergy, and the household itself. Paul Starr describes three roughly equal spheres of practice before the late nineteenth century: the home, the physicians, the lay healers. We were one sphere of three, and popular skepticism about professional claims kept us honest, kept us striving, and kept us returning.

And the patient who could negotiate was the patient who was paying. In the eighteenth-century patronage system, the fee-payer picked his physician and argued with him. In the hospitals that followed, patients were treated free, and physicians looked to professional peers rather than to patients for reputation and pay. Mary Fissell's study of the Bristol Infirmary documents wards turning into clinical material and training grounds. The standing that got lost was standing the poor never had.

The Patient Has Disappeared Before

Nicholas Jewson called it the disappearance of the sick man. In 1770 a reasonably informed patient could argue with his physician, because they shared a theory of the body. By 1870 he couldn't, because the knowledge had moved into the hospital and then the laboratory, where he had no access and no standing.

The product line changed too. At the bedside, physicians sold prognosis and prediction, which is what patients wanted. In the hospital, the emphasis moved to diagnosis and classification. The patient risked becoming a case.

Medicine got enormously better in that trade and I wouldn't undo it. The patient paid for it, and the patients in the free wards paid most.

Two Documents That Can't See Each Other

Emanuel and colleagues deserve their argument's strongest version. Reviewing the literature since January 2024, they argue that generative systems already match or beat licensed physicians at eliciting information, generating a differential, selecting tests, prescribing guideline-concordant treatment, and managing chronic disease. In one comparison, a reasoning model put the correct diagnosis first in 60% of 377 complex real-world cases. Twenty internists managed 15.9% on a subset.

And this: where the system alone beats humans alone, adding a human to the loop degrades performance against the system by itself. Our supervision makes the answer worse.

They open by rejecting the AMA's insistence on the term augmented intelligence. I'm keeping it because it holds a critical tension.

The framework names five enduring responsibilities: trust through human connection, clinical judgment, leading the evolution of practice, stewarding technology, advancing the profession. The roadmap allocates action across patients, clinicians, developers, and payers.

You could stage a fight between these two (in fact this essay started as my attempt to do just that). I think that reading is wrong, and it makes both documents look worse than they are.

The fact is that they're operating in different registers. Emanuel measures task accuracy. The framework asserts obligation. One is an empirical claim about outputs, the other a claim about duty, and what's critical is that we explore where they touch.

Read these documents as symptoms of one moment. Emanuel counts tasks because task accuracy is what our arrangement knows how to count. The framework asserts responsibilities because that's what a profession does when its authority is moving and it can't yet say where the authority went.

Emanuel's list also has a gaping hole in it. Prognostic communication isn't there. Neither is the determination of whose goals govern when two reasonable options serve competing ends. Nothing from the constellation above appears anywhere on the list of things worth measuring. Yet another example of a physician ignoring the importance of relational work.

That absence is the argument. What we measure is what we automate, what we automate is what we pay for, and the rest becomes residue.

A Positive Account, and What It Rests On

An essay that stops here is an elegy, and elegy doesn't get us very far. So let's begin with a definition:

Physicians are clinically formed interpreters of illness who accept a distinctive obligation to integrate biology, biography, evidence, uncertainty, and values into recommendations for which they remain answerable.

I chose every word in that sentence against an easier one. Clinically formed, because years of supervised formation should produce something information retrieval can't. Interpreters, which includes diagnosis and goes past it. Illness rather than disease, because the lived disruption is what patients are actually living in. Obligation, because a profession is constituted by duties it can't decline. Integrate, which concedes in advance that the machine may compute better than any of us. Recommendations, because the patient deserves guidance. And answerable, which is the whole thing: signing an output you can't explain gets you a signature, and a signature isn't accountability.

Nothing in that sentence is exclusive to a physician on a given afternoon. Individual clinicians across several disciplines do pieces of it, and some of them do those pieces better than I do.

What isn't interchangeable is the formation. Undergraduate and graduate medical education put you through years of supervised clinical hours across every organ system and every acuity, inside a structure built to expose you repeatedly to being wrong in front of someone senior. The claim is about an architecture. Character has nothing to do with it.

It's also the only defensible basis for the authority we ask for, and it creates an obligation that can fail. If those years don't reliably produce a clinician who can integrate the whole picture and answer for the recommendation, then the years are indefensible and we should be arguing about how to rebuild them. I'd rather have that argument than the one my profession usually wants to have.

Utah

This stopped being philosophical in January.

Utah launched a pilot letting an autonomous agent built by a company called Doctronic renew prescriptions on request, covering 192 drugs for chronic conditions including hypertension, diabetes, and depression. The state agreed not to enforce its unprofessional conduct laws against the developer, provided the company held to a safety and privacy contract. Physicians validate the output at first. Then it scales.

Michelle Mello walked through the design in JAMA Health Forum. The structural problem driving it is real. Patients face physical and cost barriers to seeing a prescriber. Rural clinician shortages are worsening. Medicaid coverage and ACA premium subsidies are being rolled back. Renewal requests arrive as unreimbursed work on top of an administrative load that's already crushing people.

So the autonomous agent lands first in exactly the places where physician access got withdrawn by policy. It's a substitute for care that was defunded.

The Utah Medical Licensing Board asked for suspension in April, pending review. Physician and pharmacist organizations objected, and they may well be right on the merits. Their objection also arrived after the access problem was allowed to reach the point where a state would suspend its own conduct laws to solve it.

The patients losing standing first are the ones who had the least of it. Two and a half centuries on, same as the free wards.

What I Said in April

In April I argued that augmented intelligence won't de-skill Palliative Care, because de-skilling is a pattern-recognition problem and our core act is interpretation. The narrow claim holds. Nobody's going to atrophy your capacity to sit in a room where a family is coming apart.

I didn't explore what follows from it. I treated protection from automation and protection from devaluation as one thing. My field is the completed experiment on the difference. We built a specialty around that constellation up there, and the system converted it anyway, with no machine involved. Into a consult order. Into a billing code with a time threshold. Into a template with required fields. Into a quality measure that counts whether a conversation was documented rather than whether anyone understood anything.

A CPT code hollowed out interpretive work. It didn't need an LLM's help to do that.

So when I read that these responsibilities are enduring, I don't doubt the sincerity and I don't doubt they matter. I doubt the enduring. We ran that experiment on the smallest, most self-selected, most explicitly relational specialty in medicine, under favorable conditions, with a whole field watching. It went the way it went.

What Would Have to Change

The roadmap already allocates responsibility across four groups, so I'll use its lanes.

Clinicians define competencies and standards. Then the competency set has to include what Emanuel's five tasks left out. Prognostic communication under uncertainty, and the determination of whose goals govern, specified as measurable competencies with defined performance. Whatever we decline to specify, we're electing to leave unmeasured, and unmeasured work disappears.

Policymakers and payers align payment and regulation. Utah gives you the concrete ask: there's little public information on how the system performs and no plan for independent evaluation after deployment. Published independent post-deployment evaluation should be a standing condition of any non-enforcement agreement for autonomous clinical function. A state that suspends its own conduct laws has bought an obligation to find out what happened.

And the payment ask, which my field is making loudest. If prognostic communication (in the broadest sense) is missing from the measurement set, it'll be missing from the payment set, and as physicians are disappeared from specialist palliative teams it lands as unbilled residue on whoever is standing closest. Nurses. Social workers. Chaplains. Families. Pay for the forecast. The test that generated it already gets paid for.

We've been arguing this in Palliative Care for a decade. The argument is about to go general.

Counterarguments, and Where I Might Be Wrong

The strongest objection is that I've traded technical exceptionalism for moral exceptionalism. Physicians aren't wiser or kinder than other people, and an essay concluding that our real value was always the relationship is the oldest self-serving move in medicine. I think my version survives, because it rests on what was sold and priced, and because every item in that constellation is a skill you can teach, measure, and watch someone do badly.

Second: patients want accurate care, not a meditation on professional identity. Correct, and accuracy should win every time accuracy is the question. Accuracy doesn't settle which of two accurate options a person should take. If there is one thing I've learned in my work that I wish I could convey to everyone in the world, it's the number of times in the hospital that there are three mostly reasonable ways forward or there are only two really shitty ways. Very few things in medicine reduce to a single, clear answer. Emanuel makes a lot of hay around physicians "deviating" from algorithms and guidelines while spending very little time describing why that may be: the patient right in front of me is not a hermetically sealed board question, they are a real, messy, complex human trying to do the best they can. And often they don't fit perfectly into an algorithm's box.

Where I might be wrong:

• Jewson has been criticized for 50 years as schematic rather than evidence-based, and the shifts he described were neither uniform nor inevitable. I'm leaning on a framework with known structural weaknesses because I find it explanatory. A real historian of medicine may tell me the pattern is an artifact of the model.
• Palliative Care may be a lousy generalization case. We may have converted so easily because we were never load-bearing for institutional revenue. Cardiology is load-bearing. The conversion may run differently where the money is.
• I may be underweighting how fast the measurement set can change. If prognostic communication gets specified and paid for in the next few years, my whole residue argument weakens considerably. I'd welcome being wrong here.

The Threshold

For now, I will be keeping to the phrase "augmented intelligence" not because the AMA prefers it, but because it describes a relationship, and the relationship is what's under pressure. Emanuel's framing can only describe an output, no matter how much he wants to doom-say.

Our predecessors were let into the house holding almost nothing that worked. They were let in because they could interpret what was happening, say what was likely coming, stay after it failed, and answer for what they'd recommended.

We may soon hold the most accurate tools in the history of medicine and find the invitation into the house withdrawn. If that happens, we'll have done the hollowing ourselves, and augmented intelligence will just have arrived in time to take the blame.

I am a palliative care physician, educator, and professional strategery expert. Known for turning rounds into rants and rants into teaching points. Rounds & Rants represents my views — not those of any institution or professional membership organization where I hold a role. I don't write on their behalf and they don't vet what I publish.