Answers are grounded in this paper and cite their sources.
Artificial intelligence (AI) risk is commonly managed through frameworks that identify, assess, mitigate, and monitor model-level hazards. These frameworks are necessary, but they can understate accident dynamics that arise when AI systems are embedded in operational infrastructures. This article develops a Perrowian framework for AI risk management by applying Charles Perrow’s theory of normal accidents to contemporary AI deployments. It argues that AI risk should not be defined only by model capability, technical robustness, fairness, or formal compliance. Risk also depends on system configuration: how models connect to data pipelines, retrieval systems, tools, interfaces, organizational procedures, vendors, and downstream institutions. The article maps four streams in AI risk literature, identifies their limits for accident analysis, and uses Perrow’s concepts of interactive complexity and tight coupling to explain why some AI failures become difficult to diagnose, contain, and reverse. It then proposes a systemic AI risk management model organized around boundary mapping, complexity and coupling diagnosis, containment, observability, organizational authority, and external accountability. The article contributes to risk research by reframing AI risk management as a problem of socio-technical resilience: making AI failures local, slow, visible, reversible, and accountable.
No takes yet. Share an insight, caveat, or question.
Jonny Holmström (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: