STAMP/STPA Informed Characterization of Factors Leading to Loss of Control in AI Systems

Loss of control in AI systems is a structured problem that can be analyzed systematically: The paper applies STAMP/STPA—established safety methodologies from critical systems engineering—to AI safety, demonstrating that loss of control concerns can be broken down into identifiable causal factors rather than treated as abstract risks. This approach helps organizations move from vague worries to concrete, actionable safety considerations.

Loss of control spans a spectrum from immediate to long-term risks: The research acknowledges that loss of control pathways range from rapid scenarios (like "AI self-exfiltration") to gradual "disempowerment scenarios," suggesting that responsible AI governance needs to address both sudden failures and slower erosions of human oversight and decision-making authority.

AI safety requires viewing systems as socio-technical wholes, not just technology: By framing AI systems as control structures involving humans, organizations, and technology together, the paper emphasizes that loss of control often results from breakdowns in how these elements interact—not from AI capabilities alone. This suggests safety improvements require attention to organizational practices, human roles, and system design working in concert.

Existing safety methodologies from other industries can inform AI governance: The application of STAMP/STPA—tools refined through decades of use in aviation, nuclear, and other safety-critical fields—provides a proven framework that AI practitioners can adapt, potentially accelerating the development of practical safety measures for AI systems.