Building the Analytical Infrastructure for Governing Frontier AI Development

Artificial intelligence presents genuinely novel governance challenges. Its combination of rapid capability growth, omni-use applications, and nearly free proliferation is inconsistent with templates designed for other technologies that may have dangerous applications. Recent attempts to govern AI development have had mixed results, with many proposals vetoed, revoked, weakened, or still searching for enforcement mechanisms. The deeper problem is not political will. It is that the analytical foundations for thinking rigorously about AI governance do not yet exist. We lack even a shared vocabulary. The phrase “AI safety” means at least four different things across different communities, and as a result, summits, reports, and legislative proposals talk past each other. This paper proposes building the missing analytical infrastructure through a dedicated, independent research institution with a long-term horizon. The RAND Corporation's role in nuclear strategy is a useful analogy, not because AI is comparable to nuclear weapons, but because AI, like nuclear technology in the late 1940s, presents strategic questions that existing intellectual frameworks cannot answer. Over roughly fifteen years, RAND and affiliated researchers produced the conceptual vocabulary (deterrence, crisis stability, second-strike capability) that made nuclear governance possible. AI governance has no comparable vocabulary and no institution dedicated to building one. The proposed institution would be lean, comprising roughly fourteen resident scholars and an annual budget of approximately $12 million, closer in spirit to the Institute for Advanced Study, though much smaller, than to the “CERN for AI” proposals that envision billions in shared infrastructure. It would not need access to proprietary models or training data. Its questions are about actors, incentives, and institutions, not about what any particular frontier AI model can do. Its outputs would include assessments of whether safety frameworks actually constrain behavior, formal models of AI competition dynamics, structured scenario analyses for regulators, and post-incident analysis producing shared facts.