The generation of content on social media poses a significant challenge for data analysis and informed decision-making. This study explores the application of three generative artificial intelligence models for classifying posts related to emergency situations on Platform X (formerly Twitter). The corpus consisted of tweets labeled as Emergency or Non-emergency, which were normalized using natural language processing techniques. The CRISP-ML(Q) methodology was employed to ensure traceability and quality throughout each stage, from problem understanding to results evaluation. The evaluation metrics showed that, among the models evaluated, Gemini-2.0-Flash achieved the best overall performance with an F1-score of 0.7314; GPT-3.5-Turbo stood out for its Recall of 0.9155; whereas Mistral-small exhibited a lower Precision of 0.5350. The findings suggest that, through prompt engineering, it was possible to adapt pretrained generative models to classification tasks without additional training, providing a flexible, replicable, and adaptable solution for the automated detection of emergencies in risk management contexts.