The Visibility of Disaster: How AI Reshapes the Narrative and Governance of Global Disaster Risk

From "Recording" to "Understanding": How AI Is Reshaping the Global Data Infrastructure for Disasters

In 2015, the United Nations adopted the Sendai Framework for Disaster Risk Reduction, with one of its core commitments being to replace single-hazard thinking with a "multi-hazard, systemic approach." But ten years on, most of the data underpinning this commitment still consists of isolated "event registries." Floods, earthquakes, and the resulting losses are neatly stored in databases, yet rarely answer the more critical questions: How do disasters connect to one another? How does social vulnerability amplify shocks?

A study published in *Scientific Data* in March 2026 is now filling this gap. A European research team used large language models (LLMs) and retrieval-augmented generation (RAG) to automatically extract more than 3,000 disaster events that occurred between 2014 and 2024 from millions of news sources worldwide, generating a structured "storyline" for each event and converting them into a knowledge graph. The significance of this open dataset is not merely "faster, more" data—it is that global disaster risk research is moving from tallying "what happened" toward understanding "why it happened this way" and "what comes next."

The Overlooked Cascades: Structural Limitations of Old Data Systems

Single-hazard recording systems are relatively mature under international frameworks, but disasters in the real world rarely remain "singular." Typhoons trigger floods, floods cause power outages, and power outages give rise to secondary incidents—cascade effects have appeared repeatedly in recent climate events. Yet most national databases use a single hazard as the registration unit, leaving interactions between hazards, temporal correlations, and social vulnerability almost invisible.

The knowledge graph approach intervenes by breaking disaster events down into entities and relationships, such as "floods cause landslides" and "heatwaves exacerbate droughts." The dataset includes not only the events themselves but also the narrative structure of "storylines," effectively transforming static entries into dynamic plots. For policymakers, a graphical representation of relationships reveals the sources of risk far more readily than a table of losses.

Why Now: The Intelligent Leap of LLMs and RAG

Extracting disaster information from news is nothing new. The European Media Monitor (EMM) has long aggregated media content, but traditional NLP tools fell short when it came to accurately extracting "who, when, where, and what impact" from free text and shaping it into a unified structure.

Large language models have changed this: they can understand context, identify causal relationships, and generate summaries. Retrieval-augmented generation (RAG), meanwhile, anchors the extraction process to specific news corpora, reducing the risk of "hallucination." The researchers used this workflow to supplement events in EM-DAT and validated data quality through expert evaluation and independent annotation. This is, in essence, a microcosm of AI moving from "information retrieval" toward "knowledge construction."

Technical maturity dictates that this study could only have emerged at the present moment. After 2024, the proliferation of open-source LLMs and RAG frameworks enabled academic teams to process tens of millions of articles at relatively low cost. Knowledge graph technology has also made its way into scientific data management. Several trends converge here, forming a new "global disaster narrative infrastructure."## Whose Disasters, Whose Narratives: News Data and the Global South

The corpus of this dataset comes from EMM, a monitoring system that aggregates millions of news websites. News coverage does not equal coverage of reality; it skews toward regions with stronger media infrastructure. But compared with traditional government reports, news sources have a unique advantage in capturing informal and marginal events. Many disaster events in the Global South lack statistical reports, yet they often have local media coverage. Thus, automated news mining, to a degree, fills the "visibility deficit" in international databases.

This also introduces a new geopolitical information problem: whoever owns the models and computing power may define disaster narratives. This study's choice to open its data and code, coupled with an interactive dashboard, is a form of resistance to "black-box governance." In the long run, the democratization of disaster data will affect the allocation of international relief resources and the destination of global climate adaptation funds.

From Risk Records to Decision Theater

The ultimate significance of this study lies not in producing beautiful charts, but in pointing to an emerging future: a real-time updated global disaster knowledge graph could become the underlying operating system for international disaster reduction mechanisms. If, after a tsunami forecast is issued, the system can automatically retrieve historical events of the same kind and their cascading consequences to generate a "scenario theater," decision-making would be based not only on current information but also on structured historical wisdom.

Of course, AI-generated knowledge graphs are still imperfect. Disaster narratives differ across languages and cultures, and models may amplify biases. But as the first large-scale disaster knowledge graph dataset based on global news, it provides a verifiable benchmark for multi-hazard risk assessment. The researchers explicitly note that this dataset complements, rather than replaces, traditional resources such as the UNDRR Hazard Information Profiles (HIPs). This reflects the restraint that the scientific community should exercise.

In an era when extreme climate events have become normalized, what human society lacks is not warnings, but the ability to turn warnings into structural understanding. As large language models begin to read every report on floods and earthquakes around the world and weave them into an interconnected web of knowledge, we may be witnessing a quiet transformation in global risk governance: from knowing what happened to understanding how everything happens.