Urban resilience in the face of climate-induced natural hazards critically depends on the availability of robust, high-quality datasets. However, geoscience research on landslides is hindered by limited access to consistent textual and image-based datasets. This study introduces a novel framework that leverages Generative Pre-Trained Transformer (GPT) models to synthetically generate and analyze geoscience data, addressing critical data scarcity challenges. Using prompt-engineering techniques, we autonomously generated 115 landslide events, each described by 14 parameters including event date, location (latitude/longitude), trigger cause, size, injuries, and fatalities. Statistical analysis revealed strong correlations between injury count and fatalities (r = 0.986), while seasonal analysis highlighted that large-scale landslides occur predominantly during the summer months, with higher event concentrations in India and Thailand. The synthetic dataset achieved significant diversity, encompassing 57 distinct latitudes, 56 longitudes, 42 trigger causes, and 16 categorical landslide types, thereby demonstrating its realism and potential scalability. Furthermore, visualization outputs—including heatmaps, bar charts, and AI-generated images—illustrated GPT’s capacity to effectively communicate analytical outcomes. Implemented via Microsoft Power Automate and GPT-3.5 APIs, the framework demonstrates high reproducibility and scalability, with potential to generate hundreds of thousands of hazard records for urban risk modeling. By integrating GPT into the disaster intelligence pipeline, this research provides a replicable model for enhancing urban adaptation strategies, supporting predictive analytics, and informing proactive resilience planning. Ultimately, this study establishes GPT as a transformative tool in bridging data scarcity for geoscience, offering scalable pathways to strengthen the resilience of urban systems exposed to disaster risks.