Khosla and Emanuel put a citable target on the 'human in the loop' clinical AI model

A JAMA Perspective published on August 17th argues that autonomous medical AI could eventually deliver better cognitive care than physicians, including physicians who use AI as a decision-support tool ^[1]. That puts a named, citable challenge under the deployment model that most health-AI roadmaps and both the American Medical Association and the American College of Physicians currently assume ^[4][5].

The authors are Vinod Khosla, University of Pennsylvania bioethicist Zeke Emanuel, Penn researcher Abe Baker-Butler, and Neal Khosla, founder and CEO of the AI primary-care provider Curai Health ^[2]. Vinod Khosla co-founded Sun Microsystems before starting Khosla Ventures and has argued for years that AI will make scarce professional expertise broadly available; Neal Khosla has pursued the medical version of that thesis at Curai since 2017 ^[3]. The conflict of interest is not hidden, but it is structural: two of the four authors have a direct commercial position in the model being argued for ^[2][3].

The mechanism claim is the interesting part. The AMA uses the term "augmented intelligence" to stress AI's assistive role, and the ACP has said the technology should support rather than replace physician judgment ^[4]. The JAMA authors contend this becomes counterproductive once a system consistently exceeds unaided physicians on a given task ^[5]. In a six-post thread, Khosla wrote that human review can catch model errors but can also introduce new ones when doctors incorrectly override the system, and that people are poor judges of when to trust an algorithm, particularly after it surpasses their own performance ^[6]. Applied to history-taking, diagnosis, test selection, treatment planning, and guideline adherence, the argument is that the human in the loop is a quality tax, not a safety margin ^[7].

The evidence is simulated. This is a Perspective, not a trial, and it rests on prior work including Google's AMIE, an experimental system for diagnostic conversations ^[8]. A 2025 Nature paper compared AMIE with 20 primary-care physicians across 159 simulated cases from Canada, the UK, and India: specialists rated AMIE higher on 28 of 32 measures, patient actors on 24 of 26 ^[9]. Khosla highlighted one result where physician evaluators preferred AMIE for eliciting relevant information in 97% of cases versus 50% for doctors ^[10]. The Nature authors said the synchronous text-chat format does not reflect ordinary in-person care or standard telemedicine, and that the cases could not capture physical examination, nonverbal cues, fragmented records, local practice, or the consequences of errors involving real patients ^[11]. They said deployment would require further work on safety, reliability, privacy, bias, uncertainty, and regulatory safeguards ^[12].

Later work extends the pattern without changing its character. A May 2026 study of a multimodal AMIE reading photographs, ECGs, and clinical documents across 105 simulated telehealth cases had specialists rating it ahead of primary-care doctors on 29 of 32 measures ^[13]. A disease-management study found it could reason across visits and align plans with guidelines, with the authors again flagging that real-world translation needs more research ^[14]. Notably, Google's own guardrailed experiment kept doctors reviewing generated notes and responsible for the patient-facing message, and Google said the results should not be read as proof of superiority because the workflow was designed around the AI's characteristics ^[15]. A prospective study at Beth Israel Deaconess Medical Center put AMIE into real visits as a pre-visit conversational tool ^[16].

Watch whether anyone runs the comparison the Perspective actually needs: autonomous system versus physician-plus-system, on outcomes, in real care. Every number cited so far compares AI to unaided doctors ^[9][13], which is not the copilot model anyone is deploying ^[4].