A new opinion piece in JAMA is picking a fight with the entire medical establishment, arguing that regulators shouldn’t require a human doctor to sign off on AI medical decisions. The authors—led by bioethicist and Obama healthcare architect Ezekiel Emanuel, alongside Curai Health CEO Neal Khosla—claim autonomous AI will soon outperform any doctor-machine pairing, and writing a human into the rules would lock in inferior care.
The evidence they cite is striking. In simulated patient conversations, Google’s AMIE system outscored primary care physicians on nearly every metric. ChatGPT o3 nailed the correct diagnosis first in 60 percent of 377 complex cases, while 20 internists managed just 15.9 percent. Microsoft’s diagnostic tool found correct diagnoses under budget constraints roughly four times more often than doctors, and at lower cost. The authors dismiss studies showing the opposite as outdated or methodologically weak—often because they tested older models.
The piece’s central argument hinges on a forecast: models keep improving while doctors’ skills erode through AI dependence, as a Lancet study on colonoscopies suggests. Once the machine is clearly ahead, the human checker becomes a liability. A meta-analysis of 106 experiments supports this. When AI outperforms humans, the human makes results worse by overruling the system in the wrong places. Case in point: GPT-4 alone scored 92 percent on diagnostic reasoning using real patient cases, but doctors with access to the same model scored just 76 percent.
To be fair, the authors concede real limitations. Nearly all evidence comes from simulations, not actual patient care. Physical procedures like surgery and childbirth remain firmly human territory. And autonomous systems fail in ways doctors don’t—hallucinations, internet outages, cyberattacks. But the chess parallel is hard to ignore: after Deep Blue beat Kasparov in 1997, human-machine teams dominated for years—until AI started beating the teams too, around 2017. The authors see medicine heading down the same path by 2030, which would make today’s human-in-the-loop mandates look less like patient protection and more like institutional self-preservation.
