Warnings that arrive in time: what AI early warning can deliver

An early warning earns its value through what people can do with it. A family may need time to move belongings before a river rises. A farmer needs information early enough to change planting plans. Emergency teams need a clear picture of where help is needed.

AI is contributing to each of those tasks, but the evidence takes different forms. A model can outperform a benchmark without proving that it prevented losses. A service can cover a region without reaching every resident. Reading the results carefully reveals both genuine progress and the work still required.

Flood forecasts beyond local gauge records

A 2024 Nature study tested Google's river model on locations and periods withheld from training. Its extreme-flow forecasts up to five days ahead had reliability comparable to or better than the study's zero-day GloFAS benchmark. This simulated ungauged conditions: local observations were withheld from training but used for evaluation. The result concerns river floods, with no guarantee for every location or flood type.

The paper described an operational service in more than 80 countries at that time. Coverage has since expanded: in August 2026, Google reported forecasts across 150 countries where more than two billion people live. That is the provider's coverage estimate, not a count of people who received or acted on a warning.

Research references: [1] [2]

When a forecast releases cash

In Nigeria's Kogi State, GiveDirectly paired Flood Hub data with hydrology expertise and community observations. Its account records payments to 3,250 people between August 29 and October 6, 2024, ahead of an October 21 flood peak; another 1,400 were paid on October 24 after identity checks.

This was a combined forecasting, enrollment and delivery operation. The organization reports substantial phone and in-person verification work, and warns that phone-based enrollment can exclude vulnerable people. Recipient reports of preparation and recovery are useful implementation evidence, but do not isolate the causal contribution of AI.

Research references: [3]

Monsoon guidance at national scale

In 2025, India's Ministry of Agriculture and Farmers' Welfare distributed monsoon-onset forecasts by SMS to approximately 38 million farmers. The University of Chicago-led system combined Google's NeuralGCM, ECMWF's AIFS and historical information. The collaborators reported forecasts extending up to four weeks ahead and capturing the season's unusual interruption in monsoon progression.

The timing needs care. A March 2026 preprint evaluates the blended forecasts and explains that some messages went out after the study's rainfall-based agricultural onset had already occurred. Meteorological declarations and the agricultural definition differed. “Up to four weeks” therefore describes the forecast horizon, not the lead time received by every farmer.

Separate research on the value of advance forecasts should not be presented as an income result from this mass rollout. The 2026 technical papers strengthen the evidence about forecast skill; they do not establish a harvest or income effect across all 38 million recipients.

Research references: [4] [5] [6]

Lower-cost computation, with important dependencies

ECMWF put AIFS Single into operational service on February 25, 2025. At launch, it reported improvements of up to 20% on measures including tropical-cyclone tracks and roughly a thousand-fold reduction in energy used to generate a forecast. Those are the centre's stated comparisons for that system, not an independently measured thousand-fold reduction in the full observation, training and delivery chain. AIFS uses initial conditions built from observations and conventional forecasting infrastructure.

The service has continued to evolve. An ensemble version became operational in July 2025, and both single and ensemble models were upgraded to version 2 in May 2026.

Research comparisons also show promise. GenCast outperformed the tested ECMWF ensemble on 97.2% of 1,320 variable-and-lead-time combinations in an evaluation using 2019 data. That percentage is the share of tested targets it won, not a general accuracy rate.

Aurora's 2025 Nature paper reported stronger cyclone-track performance than seven forecasting centres in its retrospective 2022–2023 evaluation, with forecasts extending to five days. These findings support further use and testing; they do not establish superiority for every future storm.

Research references: [7] [8] [9] [10]

Detection is another kind of early warning

Android's earthquake system detects earthquakes already under way and attempts to warn people before shaking reaches them. In the study published in Science in 2025, it delivered alerts in 98 countries, averaging about 18 million alerts a month over the reported operation period.

Among respondents to its timing question, 36% reported receiving the alert before shaking, 28% during, 23% after, and 13% did not know. These are self-reports from survey participants, not instrumented lead-time measurements for all recipients. They show why timing and user experience deserve attention alongside detection performance.

Wildfire detection has also moved beyond a launch announcement. FireSat's demonstration satellite launched in March 2025. Three more satellites reached orbit in July 2026, and Earth Fire Alliance released their first active-fire imagery in September.

The September release still described those three satellites as commissioning. It anticipated at least twice-daily operational data for Early Adopters by the end of 2026, hourly global revisits by 2029, and a 20-minute revisit in the early 2030s. The alliance reports sensitivity to fires as small as five by five metres. That specification and sample imagery should not be confused with a fully delivered, globally validated 20-minute warning service.

Research references: [11] [12] [13] [14]

After the event

AI also supports disaster assessment. SKAI, developed through collaboration including Google Research and WFP's Innovation Accelerator, uses imagery to identify damaged buildings. This belongs to response after a disaster. Its outputs can support assessment, but building classifications alone do not establish faster assistance or lives saved.

Research references: [15]

What builders should measure

These examples suggest a practical test for an early-warning project. Can it document the hazard, the forecast or detection time, the message's arrival, the action taken and the outcome?

That requires more than a single accuracy score. It means checking false alarms and missed events, validating performance locally, explaining uncertainty, and finding ways to reach people outside the main digital channel. It also means distinguishing a promising model from a supported operational service.

For funders and builders, the next step is to pair technical work with a named delivery partner and a measurable response plan. Ask who receives the information, what they can realistically do, and how the project will learn whether it helped. The strongest case for AI early warning will come from that entire chain.