Sovereign AI and Epistemic Resilience: Governing National Algorithmic Infrastructures

Abstract

This paper examines "sovereign AI" as nationally controlled large language models and domestic algorithmic infrastructure through the lens of epistemic resilience, arguing that reducing external technological dependence may simultaneously institutionalize domestic epistemic biases and produce new forms of epistemic closure. Drawing on structuration theory and resilience theory, AI systems are conceptualized as recursive infrastructures that codify norms of relevance and credibility, reinforcing dominant narratives through feedback loops between training, deployment, and institutional uptake. The paper identifies a dual risk structure: external epistemic vulnerability stemming from reliance on foreign AI, and internal homogenization as sovereign systems narrow domestic epistemic diversity. Using a qualitative multi-case comparison of AI governance in the EU, US, China, and smaller powers, the paper builds a framework of epistemic resilience as the capacity to detect, counterbalance, and pluralize algorithmic bias, deriving three governance implications: plural model architectures, cross-border epistemic interoperability, and institutionalized reflexivity through independent audit mechanisms. It concludes that sovereign AI governance should shift from a narrow focus on strategic autonomy to the preservation of pluralism and democratic cognition within national epistemic ecosystems.