Should AI Be Allowed to Make Medical Decisions?

“FDA-approved AI diagnostic tools already operate autonomously for specific conditions. IDx-DR reads retinal scans without physician review, and the FDA's 2018 authorization was grounded in clinical evidence that autonomous AI was safer than the available alternative — which was no screening at all. Extending this model to other well-validated conditions follows the same logic already applied.”

“Studies in Nature Medicine and The Lancet have shown AI systems outperforming average board-certified specialists in targeted tasks: detecting breast cancer on mammograms, identifying diabetic eye disease, and classifying skin lesions. The relevant comparison is not AI versus the world's best physician but AI versus the physician actually available — which in most of the world means no physician at all.”

“AI medical systems fail in ways physician errors do not. A physician who misdiagnoses makes an individual error; an AI that misdiagnoses due to training data bias produces the same error for every patient with similar demographic characteristics. When that bias favors lighter skin tones, as documented in dermatology AI, the error is structural and discriminatory, and it scales.”

“No existing legal framework assigns liability when an autonomous AI kills a patient. The device manufacturer disclaims clinical responsibility, the hospital claims the device is approved, and the patient has no physician to hold accountable. Medicine requires accountability, and accountability requires a human decision-maker present at the point of consequence.”