AI in Climate Crisis Management: Forecasts, Warnings, Limits

Extreme weather is one of those subjects where good news is rare and useful news is precious. So it's worth looking carefully at one of the more concrete developments of recent years: artificial intelligence is moving from research labs into the day-to-day machinery of weather forecasting and disaster warning. The most reliable overview I found comes from the World Meteorological Organization (WMO), the UN body that coordinates national weather services. Its page on artificial intelligence, last updated in October 2026, is notably measured: enthusiastic about the potential, blunt about the limits.

An earlier version of this article, published in early 2025, painted a broad picture of AI as a "beacon of hope" without much evidence behind it. This rewrite sticks to what can be supported, and it is a more interesting story because of that.

What AI is already doing in forecasting

According to the WMO, AI is changing how weather, climate, water and environmental information is produced and used. In plain terms, AI systems learn patterns from huge amounts of data, such as satellite and radar images, river and surface observations, and historical records, and then use those patterns to produce predictions.

This is no longer only experimental. The WMO says AI-based and hybrid systems are increasingly used in daily and medium-range forecasts. It cites the European Centre for Medium-Range Weather Forecasts (ECMWF) and its Artificial Intelligence Forecasting System, as well as systems built by national meteorological services in Germany, Canada and the United States.

The WMO also lists other uses: improving tropical cyclone track prediction, supporting "nowcasting" (very short-term forecasts for fast-developing hazards like heavy rain, lightning and strong winds), and helping with flood forecasting, impact assessment and communication. It is careful to add that many applications in nowcasting and sub-seasonal or seasonal prediction are still being tested, evaluated or piloted. That qualifier matters: "AI forecasting" is not one finished product, but a spread of tools at very different stages of maturity.

Why cheaper forecasts could matter most

The part of the WMO's account I find most compelling is about economics rather than accuracy. The organization explains that once trained, AI weather models can generate forecasts rapidly and often need much less computing power than traditional numerical weather prediction, the physics-based simulations that have long powered forecasts.

Training and maintaining advanced systems still demands data, computing infrastructure and specialised expertise, so this is not free. But the cost of producing each forecast can fall sharply. The WMO suggests this could make advanced prediction more accessible to national meteorological and hydrological services that cannot run large traditional modelling systems, and could allow more frequent updates, finer-resolution information and services tailored to local needs.

My own reading: climate risk is not evenly distributed, and neither is forecasting capacity. If cheaper models let a small national service produce decent forecasts for its own territory, that could be a bigger win for safety than a marginal accuracy gain in a country that already has excellent forecasts. That is my opinion, not a claim from the WMO, but it follows from the logic of their argument.

Floods, storms and places without radar

The WMO gives several examples. In nowcasting, deep-learning techniques can help track fast-developing storms and predict conditions over the next minutes or hours. In hydrology, AI and machine learning can support river flood forecasting, water availability assessments and decisions about how water is shared among users.

One regional project stands out. The ANDeL project (Advancing Nowcasting with Deep Learning Techniques), endorsed by the WMO World Weather Research Programme, uses satellite data to develop short-term rainfall forecasts for West Africa, where weather radar coverage is limited. When the traditional instruments are missing, learning from satellites is a sensible workaround.

The WMO also describes a pilot study on river flood forecasting carried out with Google and the meteorological and hydrological services of the Czech Republic, Nigeria, Uruguay and Viet Nam. It explored AI and machine learning approaches under different geographical and hydroclimatic conditions, and is meant to inform how new methods could complement national forecasting capacity. And in May 2026, according to the page, the WMO began a five-year initiative supported by South Korea's climate and environment ministry to strengthen flood forecasting in selected countries, exploring approaches that include AI.

Beyond prediction: turning forecasts into action

A forecast only saves lives if someone acts on it. The WMO notes that AI has potential beyond prediction: it can help combine hazard, exposure and vulnerability data, translate forecasts into expected impacts, and assist with multilingual or more accessible warning messages. Within the UN-backed Early Warnings for All initiative, it says AI can contribute to risk knowledge, monitoring and forecasting, warning dissemination and communication, and preparedness.

There is a pilot that shows this in practice. In Malawi, a project combines Bris, a high-resolution AI forecasting model from the Norwegian Meteorological Institute, with ECMWF's Forecast-in-a-Box system, which packages the tools needed to run forecasts locally. Malawian forecasters can produce AI-based forecasts, compare them with existing ones and judge their accuracy and usefulness for early warnings. Notice the design: not "trust the machine", but "test it next to what you already use".

If you want more on how this fits into the wider picture, we have covered AI in natural disaster management before, and the broader question of AI in the fight against climate change.

The limits the WMO insists on

This is where the page earns its credibility. The WMO says AI does not remove the need for observations, physical science or traditional forecasting systems. Many AI weather models are trained on decades of observations and outputs from numerical models. Physics-based models remain important for providing training data and for checking whether AI results are scientifically plausible. AI works as part of a wider system, not a replacement for it.

The organization also states plainly that the benefits are not automatic or equally shared. AI models depend on the quality and coverage of their training data, and may perform less well in regions with few observations, during rare events, or in conditions unlike the data they learned from. Some outputs can be hard to explain or physically inconsistent. Access to data, computing and skills remains uneven.

That last point deserves a pause. The regions that most need cheap forecasting are often the ones with the thinnest observation networks, which are exactly where AI models may be less reliable. The technology could narrow the gap or quietly widen it, depending on how data and investment are shared. The WMO's emphasis on support for members with fewer resources suggests it sees the risk.

Governance: trust is part of the forecast

For warnings that protect lives and property, the WMO says speed must be matched by reliability, accountability and public trust. Human experts remain responsible for evaluating AI outputs, interpreting uncertainty and turning technical information into action, and national services should remain the authoritative source of public warnings.

On the institutional side, the page says that in October 2025 an extraordinary session of the World Meteorological Congress endorsed integrating AI into WMO's global observation, data processing and forecasting infrastructure. The Congress called for stronger collaboration among the public, private and academic sectors, requested a new strategy for the WMO Integrated Processing and Prediction System that incorporates AI, and stressed open data, open-source tools and ethical safeguards.

The WMO also describes work on verification: ways to compare AI-based, physics-based and hybrid systems on forecast skill, physical consistency, robustness and performance across regions and events, using shared benchmarks and transparent documentation. This is unglamorous, but it is what will decide whether a tool is trusted at 3 a.m. when a cyclone is approaching.

What this means for the rest of us

The honest summary is that AI is becoming a real and useful part of forecasting, with its biggest promise in making good forecasts cheaper and more widely available, and its biggest risks in uneven data, hard-to-explain outputs and over-trust. The grand "AI will solve the climate crisis" framing doesn't survive contact with the details. What does survive is something more modest and more useful: better tools, tested carefully, in the hands of the people who are responsible for warnings.

One note on sourcing: much of this comes from the WMO's own description of its work, so it reflects the organization's perspective. Independent evaluations will tell us more as the pilots mature.

So here is a question to take with you. The next time a weather alert buzzes on your phone, would you want to know whether an AI model helped produce it? And how much would you need to see about how it was tested before you trusted it? Those questions will shape how comfortable societies become with AI at the front line of climate risk.