AI as Public Infrastructure: A Critical Review of the Transition from Tool to Societal Necessity

Artificial intelligence (AI) has ceased to be a collection of discrete tools. It is now a general-purpose cognitive infrastructure—an expanding layer that mediates learning, resource allocation, decision-making, and power across economies and polities. From finance and healthcare to education, defense, and public administration, AI systems increasingly shape both productivity and the legitimacy of institutions themselves. This paper conceptualizes AI as Public Infrastructure (AIPI)—a governance framework for managing the interface between globally produced AI systems and domestic institutions. AIPI operates as a filter-translation membrane, channeling the inflow of capabilities through cultural, legal, ethical, economic, and contractual layers before embedding them in critical national workflows. Drawing on real-world governance practice, we identify three dominant styles—market-led, state-led, and hybrid—defined by how authority, accountability, and coordination are distributed across public and private actors. Using country cases, we show the conditions under which AI capabilities cross into public-infrastructure status and thus warrant infrastructure-grade obligations. For small and mid-tier states, the realistic strategic aim is governed dependence—the deliberate alignment of imported frontier systems with national priorities and assurance capacity—rather than unattainable autonomy. To operationalize this perspective, we introduce the Infrastructure Status Index (ISI) as a jurisdiction- and domain-specific metric. ISI scores a given country × sector on four dimensions—Essentiality, Embeddedness, Legitimacy, Governance—to answer two questions: (1) has the national capability crossed public-infrastructure thresholds? (2) where does governance lag adoption? Taken together, the AIPI–ISI frameworks form a national/sectoral design and oversight architecture: they translate diagnosis into clear obligations and pathways for implementation, so imported AI can be embedded purposefully and accountably.