Do not be fooled by the question mark in the title of an article published in August 2026 in JAMA. When the authors — among them medical ethics professor Ezekiel Emanuel and venture capitalist Vinod Khosla — asked "Will autonomous AI exceed AI-physicians as the best medical care?", they were being rhetorical. Their answer is an emphatic yes.
Much of the medical world had begun grudgingly accepting the premise that patients get the best treatment through a hybrid approach, doctors working in consultation with well-trained bots. This article's thesis is different: AI alone delivers the best outcomes.
The debate may look distant, but it is not. The same models are in use everywhere, and the same questions will arrive in the same form. This guide unpacks it in six parts.
1. What the paper actually claims
The authors' sentence runs: "Review of all published articles on AI in medicine since January 1, 2024, shows that medicine is rapidly approaching the transition point at which AI alone will exceed physicians and physician AI-hybrids in providing the best care at five fundamental medical tasks."
Those five tasks are:
- Taking medical histories
- Establishing a diagnosis
- Identifying what tests are needed
- Prescribing treatment
- Managing chronic diseases
What is listed, in other words, is doctoring itself. The most contested line sits here: AI is now so good, the authors argue, that clinicians should refrain from meddling, because "humans in the loop degrade AI performance."
The prediction: by 2030, in many cases, a "superior autonomous AI" will likely surpass humans using AI. The authors call that forecast, in their own words, "unsettling but seems probable."
2. Who wrote it, and with what interest
This is a heading you cannot skip while reading the paper. Lead author Emanuel chairs the Department of Medical Ethics and Health Policy at the University of Pennsylvania and is a renowned oncologist. The second name, Khosla, is an investor and one of the field's oldest advocates.
Khosla has carried this thesis for years: in 2012 he wrote a TechCrunch piece asking "Do we need doctors or algorithms?", and in 2016 he produced a 101-page treatise to the same effect. At conferences he would corner Emanuel and explain that AI would be doing 85 percent of what doctors did by 2035.
Khosla's son Neal joined the effort. Neal runs Curai Health, an online company that uses AI to treat patients, with a staff of doctors to prescribe medicine and handle more complicated issues. That conflict of interest is disclosed in the paper, along with Khosla's relevant investments and Emanuel's various grants and consultancies.
Disclosure does not dissolve the problem, but it beats concealment. What to hold in mind while reading is simple: this is not an independent observation from outside the field. It is a joint paper by people, some of whom have money inside it.
3. How Emanuel changed his mind
Emanuel rejected the idea for years. In his own account: "I said bullshit, there's no way. What doctors do is too complicated."
About a year ago his friend Robert Wachter sent him the galleys of his book. Wachter, head of medicine at UCSF, sketched a world in which first-class medicine would be a collaboration between AI and doctors, while the masses in "economy class" would mostly make do with AI alone.
Emanuel began to think Khosla might be right. And a frightening question surfaced: "If AI takes over, what's left for doctors to do?"
4. The objections
One of the loudest dissenting voices is John Whyte, CEO of the American Medical Association. His objections are methodological:
- A simulation is not an experiment. Some of the studies surveyed are simulations, not blind trials with real patients.
- The findings do not all point one way. Not every study reviewed supports the article's conclusion.
- Real patients cannot make themselves understood. A February 2026 article in Nature found that in real-life cases, most patients were unable to converse effectively with large language models to access their expertise.
That third point matters most. A model reaching the right diagnosis on an exam question is not the same as the person in front of it being able to describe the complaint. The gap between laboratory performance and clinic performance sits exactly there.
Emanuel's counter is equally plain: "It's been less than four years since ChatGPT was introduced. We're making a prediction for four years from now. You don't think that AI is going to be way more advanced than a doctor?"
5. The concession from the other side
The most striking thing is the reply from Wachter, singled out in the paper's first paragraph as the man who wrongly argues AI-only treatment will be "economy class." Wachter concedes a point: "The argument they're making is important." AI is good, he says, and right now AI and humans together are better. "But that can no longer be chiseled into a tablet. There will be times when humans will muck up the performance."
He is insistent, though, that the human attributes of trained physicians are vital and irreplaceable. AI may never be as effective as a person at breaking the news of a grim prognosis, and a patient is more likely to be guided to the best path of treatment by a real clinician.
Wachter invokes what he calls the doorman fallacy: the fear that people who opened doors in apartment buildings would be out of work once doors could open automatically. In fact doormen perform myriad tasks and remain invaluable — accepting packages, feeding pets, lending a sympathetic ear. "We're going to find a version of the doorman fallacy for doctors," he says. "I can imagine a world where a doctor accepts the AI diagnosis but also performs a lot of other functions that have value."
6. The real danger: de-skilling
But what is being displaced here is not as simple as opening a door. It is hard-won expertise that comes from years of training.
A senior clinician putting residents on the spot for instant assessments is a teaching method. When AI has the answer on demand, it stops being necessary. Physicians are on a path to relying so heavily on AI that their own judgement and knowledge become less relevant.
This de-skilling process could accelerate the takeover of medicine by AI, especially as new generations of doctors stop finding it necessary to spend years learning things they can query from an always-on expert in the pocket of a lab coat.
Whyte admits this is a concern for him too: "A lot of medical schools and residency programs are debating whether physicians in their training can utilize these tools, because if you've never learned how to take medical history and do a physical, how will you ever learn?" On the other hand, there is an argument that ignoring such a compelling tool is itself a form of malpractice.
7. What the debate skips: when there is no doctor at all
Wachter's "economy class" analogy frames the comparison as human doctor versus AI. In much of the world, that is not the comparison that matters.
Where the population per physician is high, the choice is not "good doctor or AI" but "AI or nothing." If a town sees a general practitioner one day a week, a system that can follow a chronic condition is not a downgrade but an upgrade. Under those conditions the paper's thesis becomes far less contentious.
The distinction matters. The gap between a university hospital in a large city and a district hospital decides which version of the debate applies. In one the question is whether the physician's judgement should stay in the loop; in the other it is how access grows at all.
Both questions are legitimate, but they do not have the same answer. No single policy resolves both — and the debate mostly runs on the first one, because the people having it belong to the first group.
What to watch
It would be wrong to assume this debate arrives late anywhere; the models are the same models. Three things are worth tracking.
First, the kind of evidence. The question to ask of every future claim is this: simulation, or a blind trial with real patients? The difference between the two decides the whole argument.
Second, liability. The paper discusses performance but not legal responsibility. Who is accountable when a diagnosis made by AI alone turns out wrong is a question independent of technical competence, and it will be settled far more slowly.
Third, training. De-skilling is the quietest risk, because its effect shows up not today but a decade out. How medical schools treat these tools now determines what the physician of 2030 is able to do.