The “Triple identity” of generative AI in public governance: dilemmas and adaptive solutions

The profound integration of Generative Artificial Intelligence (GenAI) is reshaping the frameworks and operations of public governance. While academic interest has surged, extant literature frequently examines GenAI through a siloed lens—treating it either as a governance tool, a regulatory object, or a novel decision-making agent. Such a fragmented perspective obscures the micro-level governance paradoxes and systemic risks that emerge when these roles intersect. To bridge this theoretical gap, this study conceptualizes the public management system as a complex adaptive system, identifying GenAI as an “adaptive agent” within this framework to explore the micro-paradoxes, macro-risk emergence, and governance models driven by its triple identity. The research uses a single-case study of Singapore’s “Smart Nation” initiative. Tracking governance strategies and outcomes across various situations yields general insights for GenAI governance. The findings show a “Governance Impossible Triangle” in GenAI’s triple identity. This triangle has persistent tensions among maximizing efficiency, safeguarding the legitimacy of decisions, and risk control. To resolve these tensions, the study proposes a “Contextual Identity Configuration Model” driven by problem complexity, risk severity, value conflicts, and explainability requirements. By analyzing four prototypical scenarios—routine optimization, professional assistance, high-stakes discretion, and complex crises—the paper elucidates how management systems dynamically recalibrate algorithmic tool-intensity and agentic power through continuous feedback loops. The study concludes that policymakers must move beyond “one-size-fits-all” regulatory approaches toward an agile, context-based governance paradigm. By implementing multi-layered defense systems and joint accountability networks, governments can strive to achieve a profound alignment between technological empowerment and core public values.