An FDA for AI? Pitfalls and Plausibility of Approval Regulation for Frontier Artificial Intelligence

Defining "the product" is fundamentally harder for AI than drugs: Unlike pharmaceuticals with clear chemical compositions, frontier AI models are difficult to define as discrete products, making it unclear what exactly would be licensed or approved—a challenge the paper identifies as a key limitation of FDA-style regulation.

We face "Knightian uncertainty" about AI harms: The paper distinguishes between calculable risks and deep ambiguity, noting that we cannot reliably predict or quantify potential harms from frontier AI systems in the way the FDA can for medical treatments, undermining the scientific basis for approval decisions.

AI risks can spread across the entire ecosystem: Unlike a defective drug affecting only its users, the paper emphasizes that risks from frontier AI models are "potentially transmissible" across many actors and applications, meaning approval of a single model doesn't contain the risk the way pharmaceutical approval does.

Many actors share responsibility in ways that complicate centralized approval: The paper highlights that the distributed nature of the AI lifecycle—involving developers, deployers, users, and others—makes it difficult to apply approval regulation designed for single manufacturers, suggesting that FDA-style licensing alone cannot address the full chain of responsibility.

Better testing methods and learning from other AI regulations offer a path forward: Rather than abandoning the idea entirely, the paper suggests that improvements in evaluation techniques and insights from existing AI regulations could help overcome current limitations and make approval-style approaches more workable.