A provocative new JAMA Perspective asks whether autonomous AI could eventually outperform not only physicians working alone, but physicians working with AI.
That is a much more disruptive question than whether AI can help doctors document faster or identify something on a scan. The authors challenge one of the assumptions that has shaped much of the conversation so far: that the best version of future healthcare will necessarily be a physician with AI at their side.
There are very good reasons not to jump from that argument to “AI can replace doctors.” A great deal of the evidence still comes from simulations and bounded tasks rather than the messy reality of caring for an actual person. Medicine includes physical examination, uncertainty, accountability, conflicting priorities, procedures, longitudinal knowledge and situations in which the question itself is not obvious. The regulatory and liability infrastructure for autonomous clinical AI is also nowhere near settled.
But suppose the underlying premise eventually proves true for some parts of medicine. Suppose AI becomes exceptionally good at taking a history, generating a differential diagnosis, choosing appropriate testing, managing a straightforward chronic condition or answering many of the clinical questions that currently require physician time.
There is an obvious upside. Care could become cheaper. More people could get answers. Rural and underserved communities could gain access to clinical capabilities that are scarce today. Physicians could stop spending their attention on work that technology can perform safely and consistently. All of that is worth pursuing.
But I think there is another question sitting underneath the performance debate that healthcare needs to start asking now: What happens to the human relationship when human attention becomes the expensive part of care?
We already have a version of this problem. Time with a physician is scarce. Continuity is scarce. The ability to speak with someone who knows your history, remembers what happened six months ago and has enough time to understand why this decision is different for you is scarce. Some healthcare models have emerged specifically because patients and physicians wanted those things back.
AI could relieve that scarcity. It could also formalize it.
Imagine a healthcare system in which an autonomous system can safely handle a large portion of routine care for very little marginal cost. For many encounters, that may be completely appropriate. Some patients may even prefer it.
But physician time would then become relatively more expensive. The economic temptation is obvious: give everyone the scalable layer, and reserve significant human involvement for the cases that clinically require it or for the people willing and able to pay for more of it.
That is the point where “AI expands access” and “human relationship becomes a premium service” can both be true at the same time. And I am not sure healthcare has grappled with that tension.
Recent JAMA Network Open research on consumer perspectives suggests that patients are not simply deciding whether they like or dislike medical AI. Their acceptance is conditional and depends heavily on context.
In that study, consumers considered three broad things when deciding whether AI belonged in healthcare: how well it performed, how responsibly it was governed and whether the clinician-patient relationship was preserved. Importantly, the balance changed depending on the situation. In short-term care, performance reliability was prioritized. In longer-term care, relational engagement and structural support mattered more.
That makes intuitive sense. If I am having a stroke, I am unlikely to object because the fastest, most accurate system in the hospital is not sufficiently warm. If I have been managing a complicated chronic condition for eight years, choosing between treatments with very different consequences for my life, the calculus changes.
“Human versus AI” is probably the wrong unit of analysis. Different parts of medicine require different amounts and different kinds of human involvement.
The danger is letting cost decide where that line gets drawn before we have decided where relationship actually contributes to good care.
Because relationship in medicine is not only emotional comfort. A clinician who knows a patient across time may recognize that the usual recommendation will not work in this household. A patient may disclose something after three visits that they would never have shared during the first one. A physician may understand that the technically optimal option is not the realistic one, or that a patient asking the same question for the third time is not asking for more information at all.
Those are not arguments that AI can never acquire useful context or communicate empathetically. It would be shortsighted to assume the technology will remain where it is today. They are arguments that healthcare contains decisions whose value is partly relational, and that we should understand that value before we accidentally price it out of ordinary care.
This is also why I do not think the answer is simply “keep a physician in the loop.” That sounds reassuring, but it can become ceremonial. A physician technically reviewing an AI recommendation for thirty seconds is human involvement. It is not necessarily a relationship.
The more useful question is what type of human involvement improves the particular care being delivered. Sometimes the answer may genuinely be none. Sometimes it may be escalation when the AI encounters uncertainty. Sometimes it may be a physician interpreting a recommendation. And sometimes it may be a person who has known the patient for five years sitting down and saying, “I know what the guideline says. Let’s talk about whether that makes sense for you.”
Those are very different products. Healthcare should not pretend they are interchangeable just because each one can be described as “AI-enabled care.”
There is an interesting lesson here from Direct Primary Care and other relationship-based models. Their growth should not be interpreted simply as proof that patients will pay for better access. Many patients and physicians are paying to reconstruct things that used to be ordinary characteristics of primary care: time, continuity, responsiveness and an ongoing relationship.
That is not an argument against those models. If anything, their appeal tells us how valuable those things become when the larger system makes them scarce.
But AI creates the possibility of making the divide much sharper.
The inexpensive tier could become extraordinarily capable. That is good.
The expensive tier could become the place where you get extraordinarily capable care plus a human who knows you.
That is where I become uncomfortable.
Not because every medical encounter requires a relationship. It does not. Not because autonomous AI should be rejected if it proves safe and effective. It should not. But because the ability to be known by the person caring for you should not quietly become something healthcare reserves for people who can afford an upgrade.
The JAMA Perspective describes a current view of AI-only care as potentially becoming medicine’s “economy class,” while AI-assisted physician care occupies the higher tier. Its authors challenge that hierarchy on performance grounds, arguing that autonomous AI could eventually outperform the hybrid model in some cognitive tasks.
I think there is another reason to challenge the metaphor.
Healthcare should be extremely careful about creating an economy class and a first class around human presence in the first place.
If AI allows us to deliver good care to more people, that is a remarkable achievement. If it also gives physicians back time that administrative burden and impossible schedules have taken away, even better.
The opportunity should be to use technology to make the parts of medicine that require human attention more available, not to turn that attention into the most expensive feature in the system.
The future question is not simply whether AI is good enough to practice medicine. It is what we choose to do with human care once it is.
SOURCES & EVIDENCE
Primary evidence
JAMA: Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care? (2026)
JAMA Network Open: Consumer Perspectives on Trust in and Benefits of Artificial Intelligence in Health Care (2026)
Evidence boundary
The JAMA Perspective is a forward-looking argument rather than evidence that autonomous AI already outperforms physicians in real-world care. The JAMA Network Open study supports the claim that consumer acceptance of AI varies by context and includes preservation of the clinician-patient relationship; the article’s premium-tier argument is the author’s interpretation.
When a piece rests on my own data, I say so. When it rests on someone else’s, I say whose, and whether they funded it.