The Independent Institute article argues that AI could significantly improve medicine and drug regulation, but that government agencies may struggle to use the technology effectively. The author points to AI's potential in areas such as detecting tumors, designing antibiotics, tailoring cancer treatments and forecasting disease outbreaks. The article also cites calls for the FDA to use AI to accelerate drug approvals and reduce reliance on animal testing.
The main criticism concerns the FDA's own deployment of AI. The agency launched Elsa, an AI assistant intended to help review clinical protocols and accelerate regulatory work. According to the article, an earlier version of Elsa produced hallucinations and mischaracterized clinical-trial findings. The FDA subsequently released Elsa 4.0 in May 2026 and integrated it with a data platform called HALO, but the author argues that reviewers still report problems including fabricated studies and citations.
The article's central argument is that AI performance depends heavily on the institution deploying it. The author contrasts the FDA with private healthcare and pharmaceutical companies, where AI is being used to screen compounds, personalize treatments and assist patients. Private companies have stronger financial incentives to identify and correct unreliable systems because mistakes can result in lost customers, money and reputation. The author argues that government agencies face weaker incentives to respond quickly when an AI system performs badly.
The broader concern is particularly important in high-stakes regulation. A hallucination in a consumer chatbot may simply waste someone's time, but an incorrect AI-generated analysis inside a regulatory agency could potentially influence whether a medicine or medical product is approved or rejected. The article therefore argues that expanding government use of AI requires strong verification, accountability and incentives for correcting failures. Its broader message is not that AI should be avoided, but that giving powerful AI systems to institutions without equally strong mechanisms for checking their output can produce greater confidence without greater accuracy.
