A medical paper has pushed the debate over artificial intelligence beyond whether doctors should use it. Its authors argue that autonomous AI could outperform physicians and even human-AI teams at five core clinical tasks by 2030.
The claim covers taking histories, making diagnoses, choosing tests, prescribing treatment and managing chronic disease. It is a projection based on a review of recent research, not the result of a single clinical trial proving that autonomous systems already deliver safer care across real hospitals.
That distinction drives a sharp disagreement among clinicians about evidence, patient communication and what medical expertise should become.
The paper predicts an AI-only advantage
The authors reviewed publications on AI in medicine dating from January 2024. They concluded that the field is approaching a point where autonomous systems will provide better results than doctors and collaborative physician-AI models across fundamental tasks.
Lead author Ezekiel Emanuel, chair of medical ethics and health policy at the University of Pennsylvania, previously rejected the idea that AI could perform core medical work. He reconsidered after reading Robert Wachter’s argument that doctors and AI would provide premium collaborative care while many patients would receive AI-only treatment.
Emanuel worked with venture capitalist Vinod Khosla, Khosla’s son Neal and a researcher. Neal Khosla runs Curai Health, which provides online AI-assisted treatment with doctors handling prescriptions and complex cases. Wired notes that the paper disclosed this conflict along with relevant investments, grants and consulting relationships.
The authors acknowledge that predicting 2030 is unsettling. Their position depends on continued improvement over the next four years, rather than a claim that today’s systems can replace clinicians everywhere.
Doctors challenge the evidence and the handoff
American Medical Association chief executive John Whyte argues that some studies in the review are simulations rather than blinded real-world experiments and that not all support the paper’s conclusion.
He points to a 2026 Nature article finding that most patients in real cases could not communicate effectively enough with large language models to access their capabilities. The AMA supports medical AI, he said, but wants it used within a physician-governed care plan.
Wachter accepts that humans can sometimes reduce an AI system’s performance, yet rejects the idea that clinical relationships are expendable. A physician may be better equipped to deliver a grim prognosis, understand a patient’s response and guide a decision where technical accuracy is only one part of care.
These objections identify limits in both the evidence and the task definition. Excelling at diagnosis or test selection does not automatically establish that an autonomous system can manage consent, trust and communication in every clinical context.
Medical training faces a de-skilling problem
Even if doctors remain responsible for patients, increasing reliance on AI could change how they acquire and maintain judgment. Whyte said medical schools and residency programs are debating whether trainees should use systems that perform histories and assessments before they have learned those skills themselves.
Rejecting useful technology may also be harmful. The dilemma is whether clinicians can benefit from stronger tools without losing the expertise needed to detect a bad recommendation, manage an unusual case or act when the system is unavailable.
Vinod Khosla expects doctors to remain necessary in surgery and emergency medicine in the near term but sees less need for human judgment elsewhere. Neal Khosla expects regulators eventually to permit AI systems to prescribe medicine. Those are expectations from participants in the paper, not announced regulatory decisions.
The competing view is that doctors will shift toward functions machines handle poorly: empathy, difficult conversations and integrating clinical facts with a patient’s values. That role could be valuable, but it would still represent a major change from medicine built around hard-won diagnostic expertise.
The paper makes a forceful forecast, while its critics expose what the forecast cannot yet prove. Autonomous AI may become better at discrete medical tasks. Whether that produces better care without physicians depends on real patients, accountable deployment and evidence that extends beyond simulations and technical benchmarks.
