AI in Climate Change Adaptation: What Has Been Measured

Most writing on AI in climate change adaptation is a list of possibilities. Machine learning could forecast floods, could guide farmers, could help cities plan drainage. Some of that is now more than a possibility. A small number of deployments have been evaluated, at scale and in the field, and the results can be stated with numbers and named institutions.

This piece sorts the field by evidence. It sets out three deployments with measured outcomes, the pattern they share, where the evidence runs thin, and the constraint that none of the models can solve on their own: adaptation is badly underfunded.

What AI in climate change adaptation actually means

Climate policy separates two jobs. Mitigation reduces the emissions that drive warming. Adaptation reduces the harm from the warming already under way: floods that arrive sooner, monsoons that stall, droughts that last a season longer than the planting calendar assumed.

AI enters adaptation mostly as a forecasting and decision tool. It predicts a hazard, translates the prediction into a local recommendation, and in the better systems, delivers it to the person who has to act. Governments are starting to write this into their plans. A UNFCCC Technology Executive Committee paper found that by February 2024, 57 developing countries mentioned digital technologies in their Nationally Determined Contributions, and five referred to AI directly.

Five is a modest number. It is a fair measure of where the field stands: widely discussed, narrowly written into policy, and evaluated in only a handful of places. The companion questions of forecasting skill and emissions cuts belong to AI in climate change prediction and AI in climate change mitigation. This article stays with adaptation.

Three deployments with measured outcomes

Monsoon onset forecasts for 38 million Indian farmers

In the 2025 season, India's Ministry of Agriculture and Farmers' Welfare sent AI-based forecasts of the monsoon's onset to 38 million farmers, with up to four weeks' notice of the start of continuous rain. The forecast was chosen through an evaluation led by the University of Chicago's Human-Centered Weather Forecasts Initiative, with researchers from IIT Bombay, IISc Bangalore and UC Berkeley. Google's NeuralGCM and the ECMWF Artificial Intelligence Forecasting System (AIFS) performed best, and the team blended them with India Meteorological Department rainfall records.

The season was a real test. The monsoon reached southern India early, then stalled for weeks. According to Berkeley News, the model predicted the delayed onset in northeastern India, and a preliminary phone survey found that many farmers adjusted their planting as a result.

The economic claim attached to the programme should be read as an estimate, not a measurement. Michael Kremer, the Nobel laureate who co-directs the initiative, said dissemination likely generates more than $100 for farmers for each dollar the government invests. That figure is his projection. The 38 million reach and the correct call on a stalled monsoon are the parts that have been observed.

Flood warnings in ungauged river basins

Flood forecasting has long depended on river gauges, and the places with the fewest gauges tend to be the most exposed. A 2024 paper in Nature reported that an AI model predicted extreme riverine floods in ungauged basins at up to five days' lead time, with reliability similar to or better than same-day nowcasts from the Copernicus Global Flood Awareness System. The model runs in an operational system with free, public, real-time forecasts in over 80 countries.

The paper also sets out why lead time matters. Early warning systems reduce flood fatalities by up to 43 percent and economic costs by 35 to 50 percent, according to the studies it cites. Almost 90 percent of the 1.8 billion people vulnerable to flood risk live in low and middle income countries, which is precisely where gauge networks are sparse.

Voice advisories for dryland farmers in Senegal

The third case is smaller and more rigorous in design. The Intelligent Agricultural Systems Advisory Tool (iSAT) delivered climate-informed advice by interactive voice response, in local languages, to over 2,700 farmers in 18 Senegalese villages across the 2022 and 2023 seasons. A study in Scientific Reports compared them with matched control villages. Millet yields were 41 percent higher and groundnut yields 21 percent higher among advisory users, and their input and labour costs were 24 percent lower.

Cowpea showed no significant difference. The authors attribute that to the crop's low management needs and the small number of users growing it. The null result is worth keeping: it shows the tool was measured honestly, and that the benefit depends on how much a farmer's decisions actually move the outcome.

What these three have in common

None of the three succeeded because of a model alone. Each paired a forecast with a specific decision (when to plant, when to evacuate, what to apply) and a delivery channel that reached people where they were: a text to a phone, a public flood map, a voice call in the farmer's own language.

That matches the argument in a 2025 Nature Communications perspective on AI early warning systems. The authors call for user-centred interfaces and community feedback, for causal models that avoid spurious predictions, and for principles of fairness, accountability, transparency, ethics and sustainability. Their point is that forecast skill is necessary and not sufficient. A warning that is accurate and unread adapts nothing.

For anyone funding or designing an adaptation programme, the lesson is practical. Budget for the last mile, the translation, the channel, the follow-up survey, with the same seriousness as the model.

Where the evidence is thin

Outside forecasting for farms and rivers, the evidence base is much weaker. A 2025 review in Frontiers in Climate catalogues AI uses across urban planning, drought monitoring, disaster damage assessment and ecosystem management. It is a useful map of what is being tried. Most entries describe capability rather than measured change in outcomes for people.

The data gap is structural. Columbia's National Center for Disaster Preparedness notes that the areas needing early warning most are often the most data scarce, and that flooding alone affects 19 percent of the world's population. Models trained on global patterns can fill some of that gap, as the flood work shows, but they cannot replace local observation indefinitely.

Three honest cautions follow from the cases above:

• Reach is not impact. Thirty-eight million farmers received a forecast; how many changed a decision, and by how much, is still being measured.
• Effects vary by crop, place and decision. The cowpea result is a reminder that an average does not transfer automatically.
• Evaluation is rare. Controlled comparisons like the Senegal study are the exception in this field, not the norm.

The constraint is money, not models

The best-evidenced tools in this article are cheap relative to the harm they address. The binding constraint is elsewhere. UNEP's Adaptation Gap Report 2025 estimates adaptation finance needs in developing countries at US$310 billion to US$365 billion a year by 2035. International public adaptation finance to those countries was US$26 billion in 2023, down from US$28 billion the year before. Needs are 12 to 14 times current flows.

That gap shapes which AI tools survive. A forecasting pilot funded by a single grant can prove itself in two seasons and then stop, because nobody owns the cost of running it in year three. The Indian programme was partly supported by catalytic funding from AIM for Scale; many promising projects never find that second funder. Pricing climate exposure well is one route to moving private money into adaptation, a subject taken up in how climate risk is priced with AI.

What a research and capital room should do with this

The pattern in the evidence suggests a narrow agenda. Fund evaluation alongside deployment. Pay for delivery, not only for models. And build the path from a working research pilot to a funder willing to carry it at national scale, because that path is where most adaptation tools stall.

That last step is the problem the EX Future Summit is organised around. The summit runs 18 to 20 November 2026 as a single thirty-hour broadcast between Las Palmas and Bali, with free online attendance for verified researchers and students. Its program includes a Global Financial Raise track covering venture capital, private equity, national grants and Horizon Europe, and eight hosted sessions that match universities with private partners across continents. Among the confirmed speakers is Willi Semmler of The New School for Social Research and IIASA.

It is a working room rather than a showcase. A research group with a field-tested adaptation tool and no year-three funder is exactly the kind of participant it is built for.

FAQ

What is the difference between AI for climate adaptation and AI for mitigation?

Mitigation uses AI to cut emissions, for example by optimising energy systems. Adaptation uses it to reduce harm from climate impacts that are already happening, such as forecasting the onset of rain or a river flood so that people can act in time.

Which AI weather models are used for adaptation forecasts?

In India's 2025 monsoon evaluation, Google's NeuralGCM and ECMWF's AIFS performed best and were blended with India Meteorological Department observations. Both are general weather models; the adaptation value came from turning their output into a planting decision.

Does AI flood forecasting work where there are no river gauges?

The 2024 Nature study found that it does for extreme riverine floods: five-day forecasts in ungauged basins were as reliable as same-day nowcasts from the leading global system. Reliability still improves with more hydrological data.

Are there risks in using AI for climate early warning?

Yes. Models can learn spurious correlations, perform unevenly across regions, and issue warnings that never reach the people at risk. Researchers recommend causal methods, transparent evaluation and community feedback built into the system from the start.