Autonomous AI Outperforms Doctors Using AI Assistance, Study Review Finds

The counterintuitive case for removing doctors from the loop

Two University of Pennsylvania researchers are making a provocative argument: letting artificial intelligence systems work autonomously may produce better medical outcomes than having physicians collaborate with AI tools. Their position, detailed in a STAT First Opinion essay, rests on a systematic review of studies published since January 2024 comparing autonomous AI performance against both unaided physicians and doctor-AI teams.

Ezekiel J. Emanuel and Abe Baker-Butler examined all available research on five core cognitive medical tasks: gathering patient information, generating differential diagnoses, selecting cost-efficient tests, prescribing guideline-concordant treatment, and managing chronic conditions. Their conclusion: autonomous AI will likely surpass AI-assisted physicians at some or all of these tasks by 2030, according to STAT.

What the data shows

The performance gaps documented in recent studies are substantial. Google's AMIE system outperformed physicians at eliciting complete patient histories. ChatGPT achieved 92% accuracy on differential diagnosis compared to 74% for physicians—an 18-percentage-point advantage. Microsoft's AI Diagnostic Orchestrator produced correct final diagnoses 4.02 times more frequently than doctors while reducing average testing costs by 19.1%.

In guideline-concordant prescribing, an autonomous system called MIRA exceeded physician performance by 35 percentage points. A Stanford diabetes management study found autonomous AI achieved stable insulin dosing in 15 days, while physicians had not reached that goal after eight weeks.

Of 13 studies comparing autonomous AI to AI-aided physicians, nine showed the autonomous systems performed better. One study found autonomous AI beat physician-AI teams at diagnosis by 21.3 percentage points.

Why it matters

This research challenges the widespread assumption that human oversight always improves AI performance in high-stakes domains. The data suggest that when AI systems reach high capability levels, physician involvement may introduce more errors than corrections. This has immediate implications for how healthcare organizations structure AI deployment and how regulators approach autonomous medical systems. If the pattern holds, the optimal role for physicians may shift dramatically—away from cognitive tasks like diagnosis and treatment planning, toward procedures requiring physical presence and human connection.

The empathy question

Critics argue physicians remain essential for the "art of medicine"—trust, compassion, and difficult conversations. But the researchers cite a 2025 analysis by Alastair Howcroft showing 13 of 15 studies found statistically higher empathy ratings for AI than human healthcare professionals. Patient actors reported feeling more at ease (97% vs. 65%) and listened to (95% vs. 72%) when interacting with Google's AMIE compared to primary care physicians.

Regulatory and liability gaps

American Medical Association CEO John Whyte and others have criticized the autonomous AI conclusion, citing inadequate licensure and liability structures. Emanuel and Baker-Butler acknowledge these gaps but argue the solution is creating appropriate frameworks, not rejecting autonomous systems. They have proposed licensing structures in JAMA and are developing comprehensive liability frameworks.

The researchers note that limited real-world autonomous AI testing stems from regulatory restrictions and physician resistance, not poor capability. One real-world study in Annals of Internal Medicine examined 461 actual patient visits where AI conducted structured online chats and generated treatment recommendations. Physicians then saw patients with access to AI recommendations but produced inferior treatment plans.

The researchers acknowledge AI cannot independently perform surgeries, deliver babies, or conduct physical procedures. But for cognitive medical tasks, they argue the data point toward autonomous systems as the superior option.

These details were first reported by STAT.