Stephanie Lewis, MD (Practicing Physician)

I am a practicing physician submitting this comment regarding the regulation of generative and increasingly autonomous artificial intelligence in medicine.

I strongly support responsible medical AI. AI has tremendous potential to assist physicians with documentation, information retrieval, clinical decision support, identification of errors, and other tasks that improve efficiency and allow more time for direct patient care. However, there must be a clear regulatory distinction between AI that augments a licensed physician and AI that replaces physician judgment.

I am particularly concerned about AI systems that independently diagnose patients, determine treatment, prescribe or manage medications, determine disposition, or perform other consequential clinical functions without meaningful physician oversight.

Overall accuracy is not sufficient to establish safety. A medical AI system could perform correctly in most cases while still making uncommon but catastrophic errors. Missing a time-sensitive diagnosis or incorrectly advising a patient that emergency evaluation is unnecessary can result in permanent disability or death. Risk assessment must consider not only how often AI is wrong, but the severity of its errors.

AI should be tested under realistic clinical conditions, including atypical presentations, incomplete or contradictory histories, medically complex patients, multiple illnesses, and vulnerable populations. Benchmark examinations and retrospective datasets cannot fully reproduce clinical medicine. Patients do not arrive as complete clinical vignettes.

I am especially concerned about agentic AI capable of autonomously planning and executing clinical actions. As autonomy increases, regulatory safeguards should increase accordingly. Systems capable of initiating or modifying treatment, prescribing medications, determining whether emergency evaluation is necessary, or making other high-consequence decisions should undergo rigorous prospective clinical validation. For high-risk decisions, meaningful review by an appropriately licensed clinician should remain required.

Post-deployment surveillance is essential. High-risk systems should be monitored for performance deterioration, unexpected behavior, significant errors, and adverse patient outcomes. Significant AI-associated adverse events should be reportable and investigated.

Accountability must also be clearly established. Physicians should not automatically bear responsibility for failures of AI systems selected, purchased, mandated, or deployed by healthcare organizations. AI developers and healthcare institutions should bear appropriate responsibility for harms attributable to their products and implementation. Physicians must retain genuine authority to override AI recommendations without institutional or financial pressure.

I am also deeply concerned about economic incentives surrounding autonomous medical AI. Healthcare organizations, insurers, and technology companies may obtain substantial financial benefits if AI allows clinical work to be performed with fewer physicians. Efficiency is a legitimate goal, but reduced labor costs must never be mistaken for evidence that replacing appropriate physician staffing is clinically safe.

AI should not be considered an adequate substitute for physician care simply because it is less expensive or performs correctly most of the time. Regulators should recognize that financial incentives may encourage deployment of autonomous AI beyond what safety evidence supports. Patient safety—not cost savings or workforce reduction—must remain the priority.

Patients deserve transparency when AI substantially participates in their diagnosis or treatment and should know whether an appropriately licensed clinician reviewed consequential recommendations or decisions.

AI can become one of medicine’s most valuable tools. The appropriate objective should be physician-led, AI-assisted healthcare—not replacement of accountable medical professionals simply because autonomous technology becomes technically possible or economically attractive.

I respectfully urge the FDA to preserve meaningful physician oversight for consequential clinical decisions; require rigorous real-world validation and post-market surveillance; establish clear accountability for AI-associated harm; protect physicians’ independent clinical judgment; and protect patients from premature substitution of autonomous AI for trained medical professionals.

The ultimate purpose of medical technology must remain the health and safety of patients. Technological advancement, cost savings, and workforce reduction should never supersede that obligation.

Respectfully submitted,

Stephanie Lewis, M.D.
Practicing Physician