Leveraging AI to enhance MHEWS

AI can play a transformative role in strengthening MHEWS in an era where climate change, environmental degradation and growing socioeconomic risks are increasing the frequency and severity of disasters worldwide. AI offers new capabilities to process vast and diverse data sets, generate more timely and localized insights, and support faster decision-making. Nevertheless, AI should augment human expertise, not replace the ecosystem of MHEWS actors that include hydrometeorological and geological experts, disaster managers, local leaders, humanitarian organizations, government representatives and community members. Within and across each pillar of MHEWS, AI provides opportunities to make MHEWS simultaneously more accurate, accessible, responsive and effective. Drawing on a recent AI for EW4All Group publication (ITU, UNDRR, WMO and IFRC, 2026), this box offers insights into how AI can be used to enhance MHEWS.

AI can be harnessed to close the major gaps in disaster risk knowledge when used to address persistent problems of fragmented and inaccessible risk data. For example, by using natural language processing (NLP) and semantic searches, AI can be used to extract, structure, link and translate risk information from multiple sources and records. An example of such use comes from Afghanistan, where the World Bank supported the development of Drought-Wise, an AI-powered conversational tool designed to let technical and non-technical users access drought information using natural language. This was part of a broader drought decision support platform intended to improve access to actionable drought information. A comparable approach is also being pursued in a different setting. At Texas A&M University , researchers at the UrbanResilience.AI Lab are developing a decision-support ‘companion’ for emergency operations centres, built on large language models. This will give emergency managers rapid conversational access to historical disaster reports and other data sources during an event (Texas A&M University, 2025).

AI can also be employed to improve outdated or incomplete exposure and vulnerability baselines through computer vision, predictive modelling, satellite imagery analysis and machine learning. Further, AI can be applied to make risk information more operationally useful for MHEWS, by automating data updates, harmonizing data sets and aligning risk information with hazard forecasts, ensuring compatibility with operational warning processes. For example, the UrbanResilience.AI Lab is using AI to combine rainfall sensor and stream gauge readings, decades of local flood records, satellite imagery, location-based mobile activity and news reports, as well as photographs and video posted by residents. Combining data sources that would be incompatible or time consuming to integrate without advanced computing power provides a more complete picture of risks and impacts (Texas A&M University, 2025). AI can also be utilized to strengthen loss and damage data tracking and learning mechanisms in data poor and low-resource countries. For example, AI can be used to rapidly analyse disaster reports, extract impact information, estimate losses, synthesize evidence from multiple sources and identify patterns in the data. In the domain of detection, observation, monitoring, analysis and forecasting of hazards, experts are already demonstrating tangible added value of AI across the value cycle. Already, AI is being used in: (1) quality control in data collection;(2) pattern recognition in severe event detection; (3) nowcasting and short-range weather forecasting; (4) mid-range weather forecasting; (5) modelling of hazards, especially hurricanes; (6) refinement of forecasts through postprocessing; and (7) providing emulators for climate modelling. There are further opportunities to use AI for hazard monitoring and forecasting. For example, AI can accelerate research-to-operation cycles and democratize access to advanced forecasting capabilities. For MHEWS, it can integrate heterogeneous data sets that will be useful for impact-based forecasting.

Concrete examples of using AI for forecasting come from Malawi and Viet Nam. In Malawi, the Department of Climate Change and Meteorological Services (DCCMS), together with the Norwegian Meteorological Institute and the European Centre for Medium-range Weather Forecasting, piloted the Bris AI-based Weather Prediction model using a ‘Forecast-in-a-Box’ approach (CREWS, 2026b). This approach allows DCCMS to generate a 10-day forecast in 19 minutes, whereas DCCMS could only generate one forecast a day using traditional forecasting tools. The system also runs on a single high-end desktop rather than needing supercomputer infrastructure, demonstrating that AI can extend advanced forecasting capacity to NMHS with limited technical infrastructure. In Viet Nam, a partnership between Google, Weathernews Inc. and the Viet Nam National Meteorological and Hydrological Administration led to the integration of AI-derived flood and cyclone-tracking forecasts into a national disaster management dashboard ahead of the 2025 typhoon season. When evaluated against traditional physical forecasting models (numerical weather prediction) across the season’s seven typhoons, the AI models showed a 45 per cent reduction in cyclone track error (Weathernews, 2025). An emerging example of an integrated, operational approach is the MAZU (Multi-hazard, Alert, Zero-gap and Universal) initiative in China, developed by the China Meteorological Administration. MAZU combines AI-enabled monitoring and forecasting with multi-hazard early warning and locally adaptable service delivery, illustrating how AI can be integrated across the early warning value chain rather than applied as a stand-alone forecasting tool.

For warning dissemination and communication, AI has the potential to make warning communication systems more targeted, accessible, responsive and people-centred, thereby increasing the likelihood that warnings are received, understood and acted upon. AI contributes across five key functions: (1) multilingual message generation and adaptation, where NLP and large language models can help translate, simplify and tailor authoritative warning information for different audiences and formats, subject to validation by the responsible warning authority; (2) targeting and personalization, where AI can combine hazard, population, vulnerability and connectivity information to help tailor warnings to different locations, population groups and communication needs; (3) communication network resilience and channel optimization, where AI can analyse network conditions and population reach to recommend or dynamically route warnings through the most effective available channels; and (4) public response, feedback and misinformation management, where AI can help analyse feedback and behavioural signals, identify information gaps and potentially detect misleading or false information. Across these applications, human oversight, authoritative messaging and consistency across channels remain essential. 

Case studies illustrate how these functions are beginning to be applied in practice. In Sudan, the DARAJA + AI Early Warning Service and Partnership is working together with the Sudan Meteorological Authority, the Sudan Red Crescent, Save the Children and the mobile operators Zain and MTN. They are using AI to address language and literacy barriers by converting official weather alerts into audio messages in local languages and distributing them through familiar and trusted channels, including WhatsApp, local radio and community networks. The system is being piloted in Sudanese Arabic, Bedawit and Dongolawi, illustrating how AI-supported translation and text-to-speech can extend the accessibility of authoritative warnings to populations that may not be reached effectively by text-based alerts alone. AI is also being applied to the challenge of rapidly assessing large volumes of public information during emergencies. In Japan, the Spectee Pro platform uses NLP, image recognition and other AI techniques to analyse social media posts, images, videos and other open data, filtering irrelevant or potentially misleading content and identifying information likely to reflect real-world events. Importantly, the system combines AI-based detection with a 24/7 human verification team before critical information is disseminated to authorities, illustrating the role of AI as a decision-support tool rather than a substitute for human validation. Many AI-enabled approaches depend on Internet connectivity, smartphones or other digital infrastructure that remains unavailable to parts of the population. There is therefore a need to combine AI-enabled targeting and adaptation with channels such as CB, SMS, radio and community networks, while ensuring representative data, human oversight and mechanisms to prevent automation from reinforcing existing inequalities.

In relation to preparedness to respond to warnings, while human judgment remains critical, AI can be employed in support of four dimensions in preparedness and anticipatory action decision-making: (1) strengthening enabling environments; (2) building local preparedness capacity; (3) connecting finance to action; and (4) strengthening stakeholder collaboration. A direct example of AI being used to connect forecasts to pre-positioned finance comes from Kogi State, Nigeria. In 2024, GiveDirectly used AI-based flood forecasts from Google‘s AI-driven FloodHub to inform pre-agreed anticipatory cash transfers to be made to more than 3,250 households, 48 to 72 hours ahead of peak flooding. Recipients reported a 90 per cent drop in food insecurity and 93 per cent felt better prepared for future disasters.

In relation to the enablers of MHEWS, AI can also be employed for policy framework analysis, identification of capacity gaps, generation of preparedness scenarios, translation of forecasts into probabilistic impact assessments, optimization of financing triggers, integration of diverse information sources and in building shared situational awareness. However, there are limits to how AI can – and should – be used. For example, it cannot replace political considerations, trust or the agency dimensions of preparedness. While AI can be a useful tool in relation to financing decisions, negotiations and coordination, these functions should not be automated. Rather, effective preparedness depends on a combination of AI-supported analysis with human oversight, local knowledge, community engagement and accountable governance. 

The transformative potential of AI lies throughout the MHEWS value chain – in the connections as well as in each of the pillars. Beyond the technology, which is continuously advancing, the success of AI depends on addressing challenges including: (1) data interoperability and system architecture, where fragmented and incompatible data sets limit end-to-end integration; (2) capacity and equity gaps, which leave many developing countries without the data, connectivity and technical resources needed to benefit from AI; (3) institutional coordination, as effective MHEWS require collaboration across multiple agencies and sectors; (4) human oversight, accountability and trust, particularly where AI influences high-stakes decisions affecting public safety; and (5) decision-centred design, ensuring that AI tools are developed to support real operational needs rather than for technological experimentation.