Four conditions for AI in weather forecasts to deliver for development

Artificial Intelligence (AI) is rapidly transforming the field of weather forecasting, promising more accurate predictions, tailored insights, and new ways to interact with complex data.  The potential solutions are endless. It can help farmers anticipate bad weather, and boost food security across the world. It can save lives by helping people anticipate extreme storms and floods.  And it can make it easier and cheaper for countries and weather agencies with limited resources to run their own model.

Yet, as with any technological revolution, the integration of AI into weather forecasting also brings risks and challenges. Here, I identify four key recommendations for countries, organizations, and businesses deploying AI solutions for weather forecasts, to make sure these innovations benefit all.

Enhance the Collection, Curation, and Diffusion of Data

AI’s effectiveness in weather forecasting is fundamentally tied to the quantity and quality of the data it consumes. While high-capacity countries sustain robust data systems, poorer countries often lack proper data collection and diffusion tools. Initiatives such as the Global Facility for Disaster Reduction and Recovery (GFDRR), the CREWS initiative, or the Systematic Observations Financing Facility (SOFF) aim at closing this gap, in part to ensure universal access to early warnings. But the data gap between poor and rich countries remains immense. As new private players enter the field with AI solutions, there is a risk that public agencies—often the only entities investing in the basic infrastructure for data collection—are further weakened, and their budget diverted.

But AI solutions could instead help close the data gap, if countries implement policies to facilitate data collection and sharing and ensure that new AI data users contribute – in kind and in cash – to the data ecosystem. Financing coming from local users will enhance the financial sustainability of weather agencies and create incentives to build data systems that are more adapted to local needs, building on cheaper data collection technologies that may be a better fit for low-income countries’ contexts.

Focus on Hybrid Solutions that Combine Conventional and AI solutions

AI models have shown impressive performance in weather forecasting, but today they cannot replace existing models and tools. Direct observation forecasts are just emerging, and the best AI models still train on data created by process-based models, such as ERA5 from the European Centre for Medium-Range Weather Forecasts. While out-of-sample extreme events are an increasing threat, the ability of AI-based approach to predict them remains a topic of research. To get the best out of available technologies in the near future, hybrid solutions – using AI solutions to improve existing models – appear as the most promising option.

Forecasts’ skill levels also depend on the sector and metric, and in such a context of high uncertainty and high-stake decisions, a combination of multiple models and human expertise remains the best approach to inform decision-making. Supporting legacy weather agencies and professional weather forecasters in adopting AI, while encouraging newcomers to share data, tools, and results, is the path to maximize the benefits of such hybrid forecasting systems.

Finally, AI models remain black boxes. Even if they outperform traditional models on forecasts, only the development of process-based models deepens our understanding of how our climate system works, which is how we achieve scientific breakthroughs beyond incremental improvements. Giving up on the development of physical models, to bet on AI solutions only, would be at the expense of long-term progress.   

Use AI to Bridge the “Last Mile” between Forecasts and their Users

The greatest challenge in weather forecasting is not to produce accurate predictions but to ensure that they are useful and that users act on them. This “last mile” is where AI holds the most promise. Particularly exciting is the possibility to use AI-based models to go beyond weather forecasts, to predict the impact of weather – for instance on electricity generation and consumption, flood risks, or agricultural yields. This is what is often referred to as “Impact-based forecast” and is key to better inform decision-making and meet user needs.

Generative AI (GenAI) and Large Language Models (LLMs) enable users to interact with forecasts in new ways—asking questions, exploring scenarios, and receiving actionable advice tailored to their needs. For example, the World Bank, initiatives like AgriLLM or AIM for scale, and dozens of private startups are piloting chatbots that provide farmers with practical guidance based on weather forecasts and other sources of agricultural information.

One important opportunity is to build on the momentum around integrated agribusiness value chains, a priority of the World Bank and its AgriConnect initiative. Or on investments in energy access, for instance building on the World Bank and AfDB’s M300 initiative, which aim at connecting 300 million people to electricity in Africa. By making weather forecasts an integrated component of these value chains, we can make sure they are better tailored to users’ needs. And we can create new business models in which the economic gains from weather forecasts help finance their production, delivering sustainable funding to the sector.

Navigate Risks and Maintain Reliability and Trust

AI can deliver remarkable performance, but AI-based approaches’ failures can be spectacular. In high-stakes domains like early warning systems, mistakes can cost lives. And because generating AI-based forecasts is cheap and easy, we can already see in some countries how the multiplication of parallel forecasts, warnings, and advisory products can create confusion and a general loss of trust in weather-related products.

This is a critical issue: forecasts and early warnings lead to actions and value creation only if they are trusted by users, and trust is a fragile thing. Isolated errors from AI-based automatic forecasts could quickly threaten the credibility of institutions and tools.

It is therefore essential to move fast and slow: quickly in deploying new solutions, but cautiously in retiring old ones. Overlapping systems allow for quality control and help catch errors before they reach end users. And human expertise should remain in the loop, providing oversight and preventing the obvious mistakes that can kill trust.

Conclusion - The Lessons from Weather Forecasts are Valid for Other Issues

Weather forecasting is one of the domains where new AI tools are deployed fast, with major successes. But realizing their potential will require careful attention to data, institutions, user engagement, and risk management.

As an early adopter of AI technologies, weather forecasting provides powerful insights into the opportunities and risks that AI technologies are bringing us. The need to bring legacy institutions in the AI revolution, to develop hybrid solutions and maintain proper risk management is important in any sector where AI disrupts current practices. Hopefully, the lessons we learn deploying AI for weather forecasts will also inform and guide decision-makers as they navigate the broader AI revolution.