Evaluating the Performance of AI in Crisis Detection: A Multi-Scenario Hindcast of Extreme Precipitation Forecasts

Artificial Intelligence Weather Prediction (AIWP) models excel in global mean-error metrics, yet their efficacy in detecting low-probability, high-impact extreme events--critical for emergency response--remains under-examined. This study evaluates three leading models (GraphCast, FuXi, and Artificial Intelligence Forecasting System (AIFS)) against satellite observations and a numerical baseline across four diverse historical crises. Using a crisis-centric evaluation framework comprising Peak Amplitude Ratio (PAR), Spatial Correlation (SC), Root Mean Square Error (RMSE), volumetric Bias, and the Symmetric Extremal Dependence Index (SEDI), preliminary results reveal a systemic intensity deficit in AIWP models. While GFS maintains a PAR above 0.65 across most scenarios, AI models underestimate peak rainfall by over 90% and exhibit significant spatial displacement. These findings suggest that inherent statistical smoothing transforms catastrophic signals into benign forecasts. Consequently, over-reliance on current AIWP models for crisis detection may yield a false sense of security, potentially exacerbating rather than mitigating emergency vulnerabilities.