Artificial intelligence applied to disasters and crises management

The European Crisis Management Laboratory (ECML) does research and development in support of EU crisis preparedness and response policy. It fosters the use of science in situational awareness, adding value to the routine context analysis and operational activities of the EU's Emergency Response Coordination Centre (ERCC) and the EU's External Action Service (EEAS), and their information products. At the same time, the analyses of the ECML are meant to learn from real emergencies to enhance the EU's situational awareness capabilities by incorporating emerging dynamics and trends via anticipatory analyses. The ECML performs scientific analyses of natural and man-made disasters, conflicts and complex crises at the global level via a daily monitoring by subject matter experts. The objective is to have salient situational awareness based on open sources and scientific models delivered timely and in the right format to operational analysts. Required reaction times can be minutes, when the impact assessment is performed via automatic systems like GDACS, the Global Disaster Alert and Coordination System, up to a few weeks for more in-depth analyses. The European Crisis Management Laboratory analyses are provided in the form of (web) systems and services, scientific reports and maps, publicly distributed or for the exclusive use of the European policy Directorates Generals and External Action Service, depending on their sensitivity. The use of Large Language Models (LLM) represents an important opportunity for sense-making in a context of large amounts of open-source information and the time-constrained environment of crisis management. In this study, we examined the use of LLMs to enhance the efficiency and effectiveness of the open source-based scientific analyses performed by the ECML. Specific emphasis was put on Retrieval Augmented Generation (RAG) techniques to collect information from selected, corporate databases with trusted information, including some developed at the JRC. It provides analysts with texts drafted from a collection of selected, trusted sources of information that best fit the scope of the application. This report evaluates the merits of this approach, including the reliability of the AI-extracted information and the benefits for further training the AI engine. Finally, the opportunities and limitations for rolling out such technology in operational environments are discussed.