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

TL;DR

This paper scrutinizes whether FDA-style approval regulation can be meaningfully applied to frontier AI by formalizing approval regulation as a testing-then-veto regime and by enumerating the core assumptions it requires. It argues that frontier AI challenges—such as ill-defined risk metrics, Knightian and radical uncertainty, diffuse and distributed risk, and uncertain industry structures—undermine the fit of traditional approval regimes. The authors advocate regulatory learning and policy experimentation, drawing on histories of biomedical regulation and emerging AI-evaluation practices to outline how workable, hybrid regulatory designs might emerge. The work highlights the importance of international coordination, standardized risk assessment, and adaptive governance as critical components for any future regulatory framework governing advanced AI models.

Abstract

Observers and practitioners of artificial intelligence (AI) have proposed an FDA-style licensing regime for the most advanced AI models, or 'frontier' models. In this paper, we explore the applicability of approval regulation -- that is, regulation of a product that combines experimental minima with government licensure conditioned partially or fully upon that experimentation -- to the regulation of frontier AI. There are a number of reasons to believe that approval regulation, simplistically applied, would be inapposite for frontier AI risks. Domains of weak fit include the difficulty of defining the regulated product, the presence of Knightian uncertainty or deep ambiguity about harms from AI, the potentially transmissible nature of risks, and distributed activities among actors involved in the AI lifecycle. We conclude by highlighting the role of policy learning and experimentation in regulatory development, describing how learning from other forms of AI regulation and improvements in evaluation and testing methods can help to overcome some of the challenges we identify.

Figures (1)

• Figure 1: Possible scenarios where regulators lack complete information about frontier AI model development or testing. Arrows show which models were trained by which labs. Dark blue icons reflect regulators having complete information, and faded blue icons reflect a lack of regulatory visibility. Purple icons represent misleading or uninformative test results.