Abstract
Introduction: This article presents an in-depth single-case study of Kazakhstan and examines the latent negative effects that may accompany the expanding use of artificial intelligence (AI) in the public sector. The central premise is that such risks arise less from isolated technical performance indicators of particular AI systems than from the broader architecture of legal regulation, institutional design, and centralized data infrastructures. Methods: The study combines normative-institutional analysis with secondary qualitative analysis. The normative component examines the Law of the Republic of Kazakhstan "On Artificial Intelligence," the National AI Platform, and related digital governance projects, including Smart Data Ukimet, the Digital Family Card, and sovereign large language models. The empirical component is based on a secondary qualitative analysis of 53 semi-structured interviews with civil servants. Auxiliary analytical summaries derived from the same interview subset were used only to support retrieval and thematic consolidation. The analysis therefore reconstructs primarily ex ante and design-level risk trajectories rather than providing a full ex post assessment of realized harms, although some enabling mechanisms are already observable in existing digital infrastructures. Results: On this basis, the article develops a map of hidden AI risks adapted to the Kazakhstani context. Five main groups of risks are distinguished: political-legal and institutional risks; data- and technology-related risks; organizational and managerial risks; explainability and opacity risks; and market and environmental risks. The findings show that, in the Kazakhstani case, these risks are institutionalized ex ante through legal constructs, the distribution of administrative powers, and established data-use practices. AI is legally framed as a "tool," while a de facto redistribution of power occurs through algorithmic and infrastructural arrangements. Centralized data infrastructures, including Smart Data Ukimet, the Digital Family Card, and the National AI Platform, may intensify algorithmic stratification and systemic bias. In addition, the mass deployment of AI agents under KPI-based governance may contribute to deskilling and to the fragmentation of responsibility among developers, public agencies, and citizens. Discussion: The article argues that the principal challenge of AI deployment in Kazakhstan's public sector lies not only in the accuracy, efficiency, or technical reliability of individual systems, but also in the institutional conditions under which these systems are embedded. The Kazakhstani case demonstrates how AI-related risks can be formed before large-scale harms become empirically visible, through regulatory definitions, centralized infrastructures, and governance models that redistribute authority while obscuring responsibility. The study contributes to debates on public-sector AI governance by emphasizing the need for anticipatory institutional safeguards, transparent accountability mechanisms, and stronger scrutiny of data infrastructures in highly centralized administrative contexts.