Automation bias introduced from the use of AI in safety and regulatory applications is a human–organizational reliability problem, rather than a failure of AI models. Evidence across aviation, medicine, and nuclear contexts shows that even highly accurate systems systematically shift human behavior: experts reduce independent verification, narrow hypothesis generation, and become less likely to detect errors when automation fails.

This presentation treats automation bias as an emergent risk factor within AI-enabled risk assessment processes. In probabilistic safety assessment and regulatory review, the relevant question is not only model accuracy, but how AI outputs alter vigilance, accountability, and error detection over time. Slow, document-intensive regulatory contexts differ from time-critical operational environments, but they introduce their own risks: habituation, gradual skill atrophy, and institutionalized over-reliance.

Drawing on cross-industry evidence and existing nuclear human-factors frameworks, the talk outlines practical design and governance strategies for AI-assisted PRA and safety review. These include structured independent verification, explicit representation of model uncertainty and limits, traceability to underlying evidence, periodic manual review exercises, and organizational accountability mechanisms that prevent diffusion of responsibility. The objective is to ensure that AI integration strengthens decision-relevant risk assessment capability without degrading the human judgment that underpins safety decisions.