The large-scale integration of renewable energy has intensified the electricity market price fluctuation and encouraged strategic offering behaviors of generation companies (GenCos), including capacity withholding. This paper systematically investigates the load shedding risk driven by both physical outage and capacity withholding. First, a data-driven multi-state reliability model of generators is proposed to quantify the available capacity of power systems. Then, a novel load shedding risk assessment and responsibility allocation framework is proposed to quantify the load shedding risk and the corresponding responsibility of GenCos. Furthermore, to address the curse of dimensionality, a novel physics-informed neural networks (PINNs)-based load shedding risk assessment method is introduced. This approach significantly reduces the computation burden while enabling dynamic load shedding risk assessment and responsibility allocation that account for strategic offering behaviors. Finally, a modified IEEE 30-bus system is developed to validate the effectiveness of the proposed approach.