Risk Propagation and Resilience Early Warning in Manufacturing Supply Chains Under Generative AI–Driven Public-Opinion Disturbances

INTRODUCTION: To address three challenges arising from the large-scale generation and cross-platform dissemination of generative AI content—namely, the difficulty of quantifying external information disturbances in manufacturing supply chains, time-varying risk-propagation relationships, and the limited credibility of early warnings—we propose a Generative-AI-Aware Risk Transmission and Resilience Prediction Network GAI-RTPNet. OBJECTIVES: The network constructs a node-level exogenous disturbance representation from semantic polarity, event relevance, propagation intensity, source credibility, content novelty, AI-generation probability, and semantic consistency. It updates the edge weights of supply relationships through disturbance-driven dynamic graph attention and jointly models risk evolution across nodes and time periods using a risk-propagation gate and temporal self-attention. METHODS: Based on a shared spatiotemporal representation, GAI-RTPNet applies multitask learning to simultaneously predict future node risk, propagation probabilities, and system resilience. It further combines heteroscedastic and model uncertainty estimates to produce four-tier trustworthy early warnings. RESULTS: Experiments on GDELT 2.0, SupplyGraph, and HC3 show that GAI-RTPNet achieves values of 0.048, 0.867, and 0.041 for Risk MAE, Propagation Accuracy, and Resilience MAE, respectively, with a mean warning lead time of 4.1 d. Its performance remains stable under noise, random missingness, and cross-scenario testing. CONCLUSION: The results support the effectiveness of explicitly incorporating generative-AI-driven public-opinion disturbances into dynamic-graph risk propagation and resilience prediction, providing a unified data-driven approach to risk identification, risk-propagation inference, and proactive warning in manufacturing supply chains.