Abstract: Calls for comprehensive regulation of artificial intelligence have intensified as the technology’s capabilities have grown. Such calls conflate two distinct regulatory objects: the domestic application of AI systems, which is already underway and broadly feasible, and the development of increasingly capable AI systems at the global frontier, the subject of this paper, where the structural constraints attenuate traditional regulatory approaches. This paper examines three structural features of AI development that define the constraint envelope within which any workable governance policy must operate: a verification problem that makes enforcement fundamentally different from previous dual-use technology regimes; a self-interest problem that will shape international compliance in predictable ways; and a beneficial applications problem that creates an unfavorable political economy for sustained restraint. Understanding these constraints does not counsel despair, but it does counsel realism — a shift to shaping AI development rather than preventing it, to strengthening resilience, and less ambitiously but more honestly, ensuring that whoever sits at the frontier does so with some transparency and accountability. The result is a menu of policy recommendations that are achievable under real-world constraints.

Expand
overlay image