Epistemic Systemic Risk: Societal Error-Correction as a Regulatory Outcome for General-Purpose AI

This extended research summary presents the conceptual and regulatory framework of the article “Epistemic Systemic Risk: Societal Error-Correction as a Regulatory Outcome for General-Purpose AI.” The study examines how increasing dependence on a small number of general-purpose AI models, shared sources, retrieval systems, and evaluation rules can transform local model errors into systemic epistemic risks across education, research, media, and public administration. It defines epistemic systemic risk as a reduction in the independence of the routes through which claims are produced, tested, challenged, and corrected, and proposes societal error-correction capacity as a regulatory objective. The framework identifies five mechanisms—source concentration, recursive contamination, institutional synchronisation, corrective displacement, and contestability failure—and develops an epistemic impact assessment based on seven indicators, including source independence, synthetic recursion, cross-model error correlation, provenance retention, correction latency, human fallback capacity, and institutional contestability. It further proposes regulatory triggers and instruments designed to preserve independent verification, provenance, contestability, and institutional correction without empowering the state to determine an official truth. The complete manuscript is currently under peer review; this Zenodo record makes available only the extended research summary.