Suppose a clinical AI system can diagnose a particular condition more accurately than the average doctor, operates around the clock, costs a fraction as much as a human provider, and has been extensively validated. Why should the law still require a human clinician to repeat or approve its work? This is increasingly the choice lawmakers are faced with as they regulate healthcare AI. State and federal regulators should allow AI to independently perform defined healthcare services when it demonstrates that it can do so as safely and as reliably as existing providers.
Policymakers understandably want to protect patients from unreliable technology. But requiring safe AI to remain permanently subordinate to a human professional confuses safety regulation with occupational protection. Americans already face a shortage of medical professionals that could be as high as 86,000 physicians by 2036, spanning both primary and specialty care. As of 2026, the federal government designated areas containing more than 108 million Americans as primary-care Health Professional Shortage Areas and estimated that more than 18,500 additional primary-care practitioners would be needed to eliminate shortages. Requiring scarce professionals to review work that technology can perform equally well or better wastes exactly the resource healthcare lacks.
Human review is not synonymous with patient safety. Doctors make mistakes, miss diagnoses, and vary substantially in performance. AI makes mistakes too. The relevant question is not whether AI is perfect, but whether it performs a particular task at least as safely and accurately as the people legally allowed to perform it.
For some defined medical tasks, AI has already crossed that threshold in controlled studies. In one Nature study, researchers tested an AI breast-cancer screening system against six radiologists reading the same mammograms. The AI outperformed each radiologist. A separate Nature study trained a deep-learning system on more than two million electrocardiograms and found that it outperformed cardiology residents at recognizing six types of ECG abnormalities. at recognizing six types of ECG abnormalities.
More general systems are beginning to show similar potential. In a 2025 Nature study using simulated text-based medical consultations, Google’s AMIE system achieved greater diagnostic accuracy than traditional primary-care physicians and was rated superior on 30 of 32 clinical-performance measures by specialist evaluators. The researchers appropriately cautioned that controlled text consultations are not the same as real-world medical practice. But the results show the potential for AI to independently handle parts of routine primary care, such as taking a medical history, developing a differential diagnosis, and recommending a management plan. Whether it should be allowed to perform those tasks independently should depend on additional research and real-world evidence for each use.
Some states are instead drawing a permanent line around human providers and insulating them from machine competition. Illinois, for example, prohibits AI from independently making therapeutic decisions or generating treatment plans without review by a licensed professional. Laws like this answer the wrong question: instead of asking whether AI can safely perform a service, they reserve the service to humans regardless of performance.
AI will not replace every physician, nor should it. Many medical decisions require capabilities that today’s systems do not possess. But that is an argument for limiting AI to functions it can demonstrate it performs safely, not for writing human involvement permanently into law.
Congress has proposed a better approach. The Healthy Technology Act of 2025, H.R. 238, would allow qualifying AI systems to prescribe drugs when authorized by a state and approved or cleared by the Food and Drug Administration (FDA). In other words, it creates a pathway for AI itself—not merely a doctor using AI—to become the prescriber. It does not allow any chatbot to start writing prescriptions. AI would still have to satisfy federal requirements and receive state authorization. But the bill recognizes that healthcare functions need not permanently be reserved for humans.
The potential payoff is much larger than making doctors more productive. Autonomous AI could create an entirely new class of healthcare competitor. It could expand capacity without waiting years to train another physician, offer routine services at lower cost, reach patients regardless of geography, and increase competition on price, availability, and quality.
If an AI system cannot meet the standard required to treat a patient, state and FDA regulators should not authorize it. But if it can meet or exceed that standard, the absence of a medical degree hanging on the wall should not automatically disqualify it. With millions of Americans already struggling to find care, policymakers should expand patients’ safe and reliable options.
Justin Leventhal is a Senior Policy Analyst for the American Consumer Institute, a nonprofit education and research organization that advocates for consumers through evidence-based analysis and data. Follow Justin on X @JustinLeventhal