Epistemic Corrigibility: Institutional Design for Epistemically Conservative AI Systems

This paper introduces the concept of epistemic corrigibility — the institutional capacity and obligation to recognize when an AI system is reproducing outdated, inadequate, or epistemically unjust knowledge structures, even when technically correct, and to revise or replace it accordingly. Building on a companion paper establishing that AI systems are structurally epistemically conservative, this paper takes the normative step that analysis left open. Democratic theoretical foundations are developed through Habermas's communicative rationality, Dewey's democratic epistemology, and Christiano's epistemic account of democratic authority. Five institutional design principles are derived: epistemic audit requirements, categorical diversity mandates, epistemic sunset provisions, mandatory human interpretive oversight, and transparency of paradigmatic commitments. These principles are situated within existing governance frameworks, particularly the EU AI Act. The paper argues that contemporary failures of AI governance are primarily failures of epistemic institutional design, and that epistemic corrigibility — understood as a normative institutional design principle — is a necessary condition of democratic legitimacy in high-stakes AI deployment.