More Alerts, Less Prevention? A Computational Study of Hazard Detection, Response Capacity and Threshold Choice in Occupational Health and Safety

AI-enabled hazard detection is usually judged by whether it finds risky conditions. Prevention requires another step: somebody must have the time, authority and competence to respond. This article evaluates how that second step alters the interpretation of predictive safety performance. It reports a factorial computational experiment derived from a doctoral study of AI in occupational health and safety (OHS). Across 54 simulated operating conditions, five alert thresholds and 200 replications per condition, a score-based detector classifies 5,000 opportunities per replication. The design varies hazardous-opportunity prevalence (1%, 5% and 15%), score separation (1, 2 and 3), drift (0 and 0.75) and nominal response probability (0.40, 0.65 and 0.90), with response capacity limited to 500 alerts. At the reference threshold of 1.5, one 5%-prevalence, separation-2, no-drift condition yields mean residual hazardous opportunities of 180.8, 138.2 and 95.7 across the three response assumptions. With the same detector and nominal 65% response, imposing drift raises mean residual opportunities from 138.2 to 185.5. A sensitivity calculation also changes the preferred threshold when the relative weight of unresolved hazards and false alerts changes. The simulation shows why sensitivity, precision, workload and verified action belong in the same OHS decision model. All numerical data are computer-generated. The study does not report a deployed machine-learning system, workplace incidents, observed worker behaviour or injury reduction. Its contribution is a transparent stress test of the management choices that connect detection to prevention.