Automatic generation control (AGC) dispatch is the key task of secondary frequency regulation for interconnected grids. To generate a high-quality dispatch solution, numerous machine learning techniques, such as reinforcement learning and graph convolutional networks (GCNs), have been developed for AGC dispatch. However, they are challenging to apply to a real-world power grid due to their weak interpretability. Hence, this work proposes a novel approach to large language model (LLM)- assisted interpretability in GCN for AGC dispatch. Firstly, the impact of input features (e.g., the total regulation command and the regulation capacities of various resources) on dispatch solutions is assessed quantitatively using Shapley additive explanations (SHAP) for global interpretability. Then, local interpretability for GCN is achieved using an LLM-assisted, model-agnostic local interpretable model-agnostic explanations (LIME), which can provide actionable insights into the model’s decision logic. SHAP shows that the top eight features drive decisions, while the rest average just 15.8% of the leading feature’s contribution. Unit outputs correlate positively with their own history and negatively with others. Swapping LIME’s linear model for a decision tree boosts multiple metrics by over 50%. Experimental results further confirm that this method not only clearly uncovers the relationships between input features and AGC dispatch outputs, but also faithfully reconstructs the GCN’s decision logic across different dispatch scenarios.