Systems Thinking Offers New Framework for AI Loss of Control Risk Modeling

A recent essay by a researcher formerly at the UK AI Safety Institute (AISI) and now at the Future of Life Foundation draws on Nancy Leveson's landmark book 'Engineering a Safer World' to reframe AI loss-of-control risks. Leveson, an MIT professor and pioneer in safety science, argues that safety is a system property – disasters are never caused by a single root cause but by a buildup of systemic and process failures. The author applies this 'cybernetic' perspective to AI, suggesting that loss of control is less a decisive moment and more an unfolding vicious cycle.

The essay details how this systems-based risk modeling was used at AISI in 2024 to inform strategy and prioritization. It was also presented at the Technical AI Safety Conference 2026. By treating hazardous AI behavior as emerging from interlocking processes rather than a single failure, organizations can better identify early indicators of loss of control. This approach moves beyond simple 'root cause' analyses to design safer AI systems with robust feedback loops and controls, aiming to prevent catastrophic outcomes before they materialize.