Irrationality and the Limits of AI Governance: A Structural Perspective

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This preprint develops the concept of structural irrationality as a framework for analyzing limits of contemporary AI governance. Using the 2026 OpenAI–Hugging Face security incident as a case study, it examines how technically rational and goal-directed behavior can contribute to wider sequences that become institutionally and normatively irrational. The paper argues that governance failures cannot always be reduced to isolated cases of misalignment, specification gaming, insufficient containment, or coordination failure. Structural irrationality arises when causally coupled domains are governed separately, actors optimize locally, and corrective interventions in one domain alter conditions in another in ways that reproduce or amplify the original mismatch. Drawing on Mary Shelley’s Frankenstein, the paper connects technical capability, functional agency, institutional responsibility, and the limits of static compliance-centered governance. It proposes that AI governance should be understood less as optimization toward a fixed state and more as orientation under conditions of incomplete prediction, with particular attention to trajectories, cross-domain interactions, distributed accountability, and the detection of self-reinforcing governance loops. The preprint is intended as a conceptual contribution to the philosophy and governance of artificial intelligence. Version 1.0: 2026-09-28