Paper

Abstract

What determines whether an AI regulatory framework functions as a coherent governance system or a collection of independent provisions? Existing comparative analyses catalogue cross-jurisdictional differences but offer limited structural explanation for why they arise. This paper introduces a 17-sub-dimension coding framework based on the People–Processes–Platforms (PPP) model and applies it to 12 regulatory instruments across 10 jurisdictions, producing a provision-level comparative matrix of global AI governance. All 170 classifications are validated against an independent AI-assisted coding pipeline (87.6% agreement, n=170), with reliability predicted by regulatory text legibility rather than document complexity. From this dataset, we identify two structural features that account for substantial variation in the observed data. First, AI system classification functions as a cross-dimensional regulatory trigger: jurisdictions with formal classification schemes (EU, China, Colorado, Canada) exhibit cascading obligation structures averaging approximately 12 mandatory provisions, while those without (UK, Singapore, Japan, OECD, UNESCO) average fewer than one. This difference is driven by architecture, not ambition. Second, we identify an institutional creation threshold that bounds global governance convergence: provisions implementable through existing institutional capacity are addressed in 76% of jurisdiction–sub-dimension pairs, while provisions requiring new governance infrastructure (certification, incident reporting, regulatory sandboxes) are addressed in only 27%. This pattern holds across all regulatory philosophies and explains why voluntary frameworks plateau at 65% coverage. Among the jurisdictions studied, these findings suggest that the most consequential barriers to harmonization may not be principled disagreements over values but structural differences in regulatory architecture and institutional creation capacity. We derive five testable propositions and maintain the coding framework as a living benchmark.