September 6, 2026 09:00 AM PST
(PenniesToSave.com) – Somewhere in California, Rhode Island, or New York this morning, a patient will check in for a primary care visit and never know that the first clinician on the case was software. The health network Akido operates a tool called ScopeAI, and according to chief people officer Jim McGee, it performs a clinical investigation and produces recommended diagnoses and care plans before a physician ever connects with the patient [5]. That is not a forecast. That is an appointment someone is keeping today.
Now medicine’s flagship journal has published an argument that goes considerably further. In a perspective piece released on August 17, 2026, a group of authors led by University of Pennsylvania bioethicist Ezekiel J. Emanuel contends that for the thinking parts of medical care, artificial intelligence working alone is likely to deliver better results than a physician alone or a physician assisted by AI [1]. Their projection is that autonomous systems will be ready for real world deployment in some, and possibly many, clinical workflows by 2030 [1].
The claim deserves a fair hearing. It also deserves the scrutiny that any argument gets when the people making it have a financial stake in the answer, and when the profession being written out of the room has objected on the record.
Quick Links
- What Does the JAMA Paper Actually Claim?
- Who Is Making the Argument, and What Do They Stand to Gain?
- Why Are Physicians and Their Own Association Pushing Back?
- What Is Already Running in Real Clinics Today?
- Who Is Responsible When the System Gets It Wrong?
- What Should Households Do Before 2030 Arrives?
What Does the JAMA Paper Actually Claim?
The paper narrows medicine down to five cognitive tasks: gathering relevant information from a patient, building a differential diagnosis, choosing which tests to order, prescribing treatment that matches published guidelines, and managing chronic disease over time [1]. On each of the five, the authors assemble published comparisons that favor the machine.
The diagnostic numbers are the ones that draw attention. Across 377 real world complex cases, ChatGPT o3 placed the correct final diagnosis first 60 percent of the time, while 20 internal medicine physicians managed that in 15.9 percent of a 302 case subset [1]. On treatment, Google’s Articulate Medical Intelligence Explorer was rated favorably on prescribing 90 percent of the time against 37 percent for licensed primary care physicians, and a Cedars-Sinai system produced optimal treatment recommendations in 77.1 percent of 461 real patient cases against 67.1 percent for physicians [1].
One cost figure matters more to a household than any accuracy percentage. Working under an $8,000 budget across 56 complex cases, Microsoft’s AI Diagnostic Orchestrator reached the correct diagnosis 4.02 times more often while spending 19.1 percent less, $2,396 against $2,963 [1]. For families already managing medical debt alongside other obligations, a $567 swing on a single workup is not an abstraction.
Then comes the turn that separates this paper from the rest of the field. The authors argue that once AI outperforms humans at a task, putting a human back in the loop makes the result worse rather than better, because the errors physicians introduce start to outnumber the ones they catch [1]. They point to chess, where machine and human teams dominated for roughly twelve years before the machines alone pulled ahead for good [1]. Writing in Wired, Steven Levy noted that the question mark in the paper’s title is rhetorical and the authors’ answer is affirmative [3].
Once the machine is better, the argument goes, the doctor checking its work becomes a source of error rather than a safeguard.
Who Is Making the Argument, and What Do They Stand to Gain?
Nothing here suggests anyone did anything improper. JAMA published every disclosure in full, which is exactly how the system is supposed to work. But readers are entitled to know who signed the argument.
Neal Khosla, a bylined co-author, is chief executive officer of Curai Health, a virtual primary care provider [1]. Chief Healthcare Executive reports that he wrote on X that Curai was founded with the mission of proving AI could outperform humans at the core cognitive work of medicine [6]. In other words, one of the paper’s authors runs a company whose stated founding purpose is the paper’s conclusion.
Vinod Khosla, who filed a conflict disclosure on the piece, is an early investor in OpenAI, Curai Health, and Limbic [1], and is reported to be leading a group seeking to purchase the Seattle Seahawks [6]. Neal Khosla also holds a patent for ensemble machine learning systems and has a patent pending on a multiagent architecture for cross checking autonomous AI reasoning pipelines [1].
Emanuel’s own disclosure runs long. It lists honoraria from more than a dozen organizations, grants including support from the Bill and Melinda Gates Foundation, advisory roles at Notable Health, FeelBetter, and Clarify Health Solutions among others, and consulting work for Korro [1].
One more detail is worth flagging because coverage has been inconsistent. Futurism described the authorship as Emanuel and Vinod Khosla [2], while the journal itself carries Emanuel, research fellow Abe Baker-Butler, and Neal Khosla on the byline with Vinod Khosla disclosed separately [1]. The journal record governs. None of this makes the underlying data wrong. It does mean the case for removing physicians from the decision was made in part by people who build and finance the systems that would take their place, and that context belongs in the reader’s hands rather than a footnote.
Why Are Physicians and Their Own Association Pushing Back?
The objection from organized medicine arrived quickly and from the top. John Whyte, chief executive officer of the American Medical Association, told Wired that many of the studies the paper surveys are simulations rather than blinded experiments, which medicine treats as the standard for this kind of claim, as reported by Futurism [2]. Whyte said the association sees potential in the tools, but that they belong inside a care plan “governed by a physician” [2].
He also pointed to research finding that most patients struggled to communicate with AI chatbots well enough to get the expertise they needed [2]. That criticism lands with unusual force because the paper concedes the same point. Its authors cite findings that the exchange of information between the model and the user is a particular failure point, and they acknowledge that feeding written case summaries directly to a model may overstate how it performs with an actual person [1].
Cardiologist Eric Topol commented publicly that the underlying studies were not drawn from real world medicine, so the outperformance claim remains unproven [6]. In the same week the paper appeared, the AMA and the Digital Medicine Society issued a framework holding that physicians remain responsible for weighing evidence, navigating uncertainty, balancing competing risks, and making decisions that reflect each patient’s circumstances, preferences, and goals [6].
Even Robert Wachter of the University of California, San Francisco, whose book the paper argues against, allows that there will be moments when human involvement degrades performance, while calling the wholesale automation of physicians a doorman fallacy, as reported by Futurism [2]. Emanuel, for his part, concedes that these systems will make mistakes and that no one should regard them as foolproof [6].
What Is Already Running in Real Clinics Today?
Set the argument aside and look at the installed base. A March 2026 AMA survey found that 81 percent of doctors, roughly four in five, report using AI at work, in areas including documenting patient conversations, locating current research, and preparing patients for discharge. That share has doubled since 2023 [6]. Futurism separately cites internal company data from the clinician tool OpenEvidence suggesting roughly two thirds of American physicians use that single product, though the company has published no methodology behind the figure [2].
The distinction that matters is scope. Almost all of that 81 percent is assistive use, not autonomous decision making. The doctor still decides.
Akido is the closest thing in view to a live test. McGee says the network runs more than 240 physicians serving half a million patients across three states, and the outlet describes it as the first AI native health system in the country, a characterization that traces back to the company itself [5]. Notably, McGee says the organization is not attempting to replace doctors but to reduce cognitive load and hand physicians better information [5]. That is precisely the physician controlled hybrid the JAMA authors argue underperforms [1].
Regulatory scale is already substantial. Zeeshan Tariq, chief digital and information officer at CooperVision, writes that the FDA’s public list now includes more than 1,500 AI enabled medical devices, with pathways being built for certain products to improve after launch within reviewed boundaries [4]. He also reports that PwC’s 2026 Global CEO Survey found 56 percent of chief executives saw neither higher revenue nor lower costs from AI, a single source figure offered without published methodology [4]. Tariq notes that adoption near diagnosis and treatment has been deliberate [4].
Four in five doctors already use AI. Almost none of them have handed over the decision.
Who Is Responsible When the System Gets It Wrong?
This is where the argument thins out, and Emanuel says so himself. He identifies three structures that would have to exist before any of this reaches scale: regulatory oversight determining when a system is ready to be used on people, liability rules determining who answers when it errs, and a payment mechanism determining how any of it gets billed. He says none of the three is currently set up [6].
That gap is the whole ballgame for a household. A patient harmed by a physician has a defined path to recourse, built over generations, with a licensed individual at the end of it. A patient harmed by an autonomous recommendation has no settled path at all right now. Accountability that cannot be attached to a named party is not accountability.
The paper concedes that failure modes differ. Tail risks from lost connectivity, cyberattacks, and hallucinated output will be greater with autonomous systems than with hybrids, and must be weighed against any accuracy gains [1]. Emanuel acknowledges that hospitals facing power outages or cyberattacks will still need physicians capable of practicing without the software [6]. He also observes that the public reacts differently to the two kinds of failure, saying people are “more forgiving of death by human than death by machine” [6].
Fairness requires noting the gap cuts both ways. Emanuel says there is no reliable measure of how often human physicians fail to diagnose correctly, only that the rate is known to be well above zero [6]. Meanwhile the paper cites evidence that exposure to AI erodes physician skill over time, a trend it describes as difficult to avoid without deliberate retraining or nonuse [1]. If the fallback atrophies while the systems are still brittle, households absorb that cost.
What Should Households Do Before 2030 Arrives?
Start by asking at the appointment. It is reasonable to ask whether AI was used in the workup and whether a physician independently reviewed the recommendation before it reached you. Under the framework the AMA and the Digital Medicine Society released, that judgment remains the physician’s responsibility [6]. At systems built like Akido, McGee describes the clinical investigation as taking place before the physician connects with the patient [5], which means part of the encounter may be complete before anyone with a license has looked at it.
Watch the billing line as closely as the accuracy claims. The paper’s own testing showed a lower diagnostic spend, $2,396 against $2,963 in one comparison [1], but whether a dollar of that reaches a family’s bill depends entirely on the payment structures Emanuel says are unresolved [6]. Building a habit of reviewing charges line by line is the practical defense while those rules get written.
Keep your own records portable. As the party generating the recommendation changes, the value of holding a complete personal history rises.
Finally, separate the categories. Publication in JAMA carries weight, and a perspective essay is an argument, not a trial result. Emanuel says medical schools will need to rethink their priorities, with more emphasis on empathy and connection with patients, a shift he calls a good thing [6]. That is a telling admission from the paper’s lead author about what the machine does not do.
Final Thoughts
What we have is a serious argument, published in a serious journal, resting on an evidence base the authors themselves describe as largely simulated and in need of real world testing [1]. Baker-Butler has said the goal is to get health system leaders to confront the possibility rather than avoid it [6]. That is a defensible aim. It is not the same as proof.
Two facts survive every disagreement across the reporting. Physician use of AI is already widespread, at 81 percent and climbing [6]. And the rules governing what happens when it fails have not been written [6]. Those two facts pointing in opposite directions is the entire problem.
There is a real conservative case for the efficiency here. Faster diagnosis, fewer unnecessary tests, and a lower workup cost are genuine goods, and a system that spends less to reach the same answer deserves a hearing rather than reflexive resistance. Households already stretched by health costs should welcome anything that honestly reduces them, and folding those savings into a household budget you actually maintain is how efficiency becomes something you can feel.
But efficiency is not the only value on the table, and it is not the highest one. The person on the exam table should be able to name who is accountable for the decision being made about their body. Right now, by the lead author’s own account, no one can. Four years is not long to answer a question that large, and the regulatory, liability, and payment decisions that will settle it are being made now, mostly outside public view. That is the part worth paying attention to.
Works Cited
[1] Emanuel, Ezekiel J., et al. “Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care?” JAMA, 17 Aug. 2026, jamanetwork.com/journals/jama/fullarticle/2852952.
[2] Landymore, Frank. “American Medical Association Fires Back Against Paper Finding AI Is Now Outperforming Doctors.” Futurism, 4 Sept. 2026, futurism.com/artificial-intelligence/american-medical-association-fires-back-ai-doctors.
[3] Levy, Steven. “AI Has Human Doctors Asking: What’s Left for Us?” Wired, 28 Aug. 2026, wired.com/story/ai-has-human-doctors-asking-whats-left-for-us.
[4] Tariq, Zeeshan. “The Future of AI in Healthcare: Moving from Hype to Business Value.” Rochester Business Journal, 4 Sept. 2026, rbj.net/2026/09/04/the-future-of-ai-in-healthcare-moving-from-hype-to-business-value.
[5] Colletta, Jen. “Adoption Lessons from the Nation’s First AI-Native Health System.” HR Executive, 2 Sept. 2026, hrexecutive.com/adoption-lessons-from-the-nations-first-ai-native-health-system.
[6] Southwick, Ron. “Ezekiel Emanuel: ‘We’re Going to Have Autonomous Clinical AI.'” Chief Healthcare Executive, 25 Aug. 2026, chiefhealthcareexecutive.com/view/ezekiel-emanuel-we-re-going-to-have-autonomous-clinical-ai.