Abstract: To address the challenges in emergency decision-making for major incidents, where relying solely on human judgment makes it difficult to integrate multi-source information and converge opinions to reach a consensus, this study explores a human-AI collaborative consensus-reaching method for group decision-making by introducing large language models (LLMs). Based on the influence of trust on group preferences, a dynamic human-AI trust measurement framework is proposed. It integrates initial trust and dynamic trust to construct a human-machine hybrid evaluation model, overcoming the limitation of existing studies that only use LLMs as assistants, and leveraging the strengths of both artificial and human intelligence. Subgroup division is carried out by combining human-AI trust levels and expert consensus status, and a differentiated opinion management mechanism is established. The overall consensus level is improved through feedback-based opinion revision and weight iteration. Finally, using the case of persistent high-temperature and drought in southern China, it is shown that compared with baseline methods, the proposed approach reduces the number of consensus-reaching iteration rounds by 33.3%, lowers the total range of expert opinion adjustments by 69.5%, and increases the matching degree of solution rankings by 13.8%. This study provides a reference for solving the human-AI collaboration challenges in emergency decision-making for major incidents.