Vinod Khosla (@vkhosla) and three co-authors argued Monday that autonomous medical AI could eventually deliver better cognitive care than physicians, including doctors who use AI as a decision-support tool.
The claim appeared in a JAMA Perspective published August 17th, written by Khosla, University of Pennsylvania bioethicist Zeke Emanuel (@ZekeEmanuel), Penn researcher Abe Baker-Butler (@AbeBakerButler), and Neal Khosla (@nealkhosla), founder and CEO of AI primary-care provider Curai Health. Vinod Khosla, who co-founded Sun Microsystems before starting Khosla Ventures, has spent years arguing that AI will make scarce professional expertise broadly available. Neal Khosla has pursued the medical version of that thesis at Curai since 2017. (jamanetwork.com)
The authors are challenging the deployment model favored by much of organized medicine. The American Medical Association uses the term "augmented intelligence" to emphasize AI's assistive role, while the American College of Physicians has said the technology should support clinical decisions rather than replace physician judgment. The JAMA authors contend that this structure could become counterproductive once an AI system consistently exceeds unaided physicians on a given task. (jamanetwork.com)
In a six-post thread, Khosla said human review can catch model errors but can also introduce new mistakes when doctors incorrectly override an AI system. People are poor at determining when to trust an algorithm, he wrote, especially after the algorithm's performance surpasses their own.
That is the Perspective's central bet: the familiar "human in the loop" safeguard may reduce quality in cognitive tasks where the human reviewer is less accurate than the system being reviewed. The authors apply that argument to history-taking, diagnosis, test selection, treatment planning, and adherence to clinical guidelines. They also argue that studies using generic language-model interfaces may understate the performance of systems built specifically for clinical reasoning.
The evidence remains largely simulated
The article is a Perspective, not a new clinical trial. Its case rests on previously published research, including Google's Articulate Medical Intelligence Explorer, or AMIE, an experimental system designed for diagnostic conversations.
A 2025 Nature paper compared AMIE with 20 primary-care physicians in a blinded study using 159 simulated case scenarios from Canada, the UK, and India. Specialist doctors rated AMIE more highly on 28 of 32 evaluation measures, while patient actors favored it on 24 of 26 measures. Khosla highlighted one result in which physician evaluators favored AMIE over doctors for eliciting relevant information from patients in 97% of cases versus 50%. (nature.com)
The experimental setup sharply limits what those numbers prove. Doctors and AMIE communicated with trained patient actors through synchronous text chat, a format the Nature authors said does not reflect ordinary in-person care or standard telemedicine. The cases could not capture physical examination, nonverbal cues, fragmented records, local clinical practices, or the consequences of errors involving real patients. The researchers said moving AMIE from a prototype into care would require additional work on safety, reliability, privacy, bias, uncertainty, and regulatory safeguards. (pmc.ncbi.nlm.nih.gov)
Google has since expanded AMIE into multimodal and longitudinal-care experiments. A May 2026 study evaluated a version that could interpret photographs, electrocardiograms, and clinical documents across 105 simulated telehealth cases. Specialists rated it ahead of primary-care doctors on 29 of 32 measures. A separate disease-management study found AMIE could reason across multiple visits and align treatment plans with guidelines, while its authors again said real-world translation required further research. (pubmed.ncbi.nlm.nih.gov)
Google's own work also leaves room for physician oversight. In an experiment involving a guardrailed version of AMIE, doctors reviewed generated notes and retained responsibility for the message delivered to the patient. Google said the results should not be interpreted as proof that the system was superior to clinicians because the workflow had been designed around the AI's characteristics. (research.google)
A prospective study at Beth Israel Deaconess Medical Center moved AMIE into real clinical visits as a pre-visit conversational tool. Patients generally rated the interaction favorably, and doctors said its transcripts helped them prepare. Patients still saw a clinician afterward, and the study did not establish that autonomous AI produced better diagnoses, treatments, or health outcomes. (research.google)
A thesis with commercial stakes
The authors' affiliations matter to the argument. Neal Khosla runs Curai, which sells virtual primary care supported by proprietary AI and machine-learning systems. Curai lists Khosla Ventures among its investors. Vinod Khosla's venture firm has made AI-delivered expertise, including health care, a core investment thesis. (jamanetwork.com)
Those stakes do not invalidate the Perspective. They explain why the authors are pushing beyond the safer consensus around physician augmentation. An autonomous system could serve patients without requiring a doctor's attention for every interaction, changing the labor and cost structure of primary care. A physician-led product remains constrained by clinical staffing, scheduling, licensing, and review capacity.
The evidence has yet to cross the line that matters most: autonomous performance on diverse, consequential cases involving real patients, measured through clinical outcomes rather than ratings of simulated conversations. Regulation, liability, patient consent, escalation rules, and accountability for an incorrect diagnosis remain implementation problems even if model performance clears that bar.
Khosla's "game mostly over" framing runs ahead of those facts. The JAMA Perspective still forces a harder question on health systems and AI founders. Human oversight cannot be treated as an automatic safety improvement. It must be tested as rigorously as the model, with evidence showing which participant should control each clinical decision.