| AI incidents increasingly span multiple risk domains simultaneously, creating compound risks that are harder to predict and manage. This study develops a machine learning framework for predicting compound risk in AI incidents using 1,350 incidents from the AI Incident Database, classified across seven risk domains using the MIT AI Risk Repository taxonomy. We compare compound risk rates and risk domain signatures across nine sectors, finding significant sector-dependent variation (χ² test, p<.001), with Finance exhibiting the highest compound risk rate (86.8%) and education occupying a distinctive mid-range position (52.1%) characterized by elevated Misinformation and Malicious Actors & Misuse domains. Gradient Boosting achieves the best classification performance (AUC=0.896, F1=0.827) for predicting compound risk, with primary risk domain and intent as the most important features. A transfer learning experiment reveals that models trained on cross-sector data experience an 8-9% F1 performance drop when applied to education incidents, indicating sector-specific risk dynamics that general models do not fully capture. Temporal analysis reveals a near-doubling of compound risk rates following the launch of ChatGPT in late 2022. These findings contribute a predictive framework for anticipating multi-domain AI risk cascades and demonstrate the value of sector-aware risk assessment in AI governance. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.