AI Future Fridays: Disaster Response at Machine Speed

Welcome to AI Future Fridays. Every week this column takes one force shaping the near future, tells you the truth about it, and hands you something to build before Monday. This week: early warning, and the strange moment we've arrived at where the machines see the storm coming and the message still doesn't always arrive.

Here's what's different about this moment. For a century, the hard part of disaster warning was knowing. Now the hard part is telling. That's a much better problem to have, and it's also the first version of this problem that an ordinary person can solve for themselves on a Saturday afternoon.

You're not a bystander in this one. You're the last mile.

The Backstory of Disaster Early Warning Systems

Every future has a lineage, and this one starts with a telegram that nobody was allowed to send. The myth worth killing right at the top is that disasters kill people because we can't see them coming. Often we could see them coming fine.

On September 8, 1900, a hurricane came ashore at Galveston, Texas and killed somewhere around 8,000 people, which still makes it the deadliest natural disaster in United States history according to the National Hurricane Center. The forecast existed. Meteorologists at the Belén Observatory in Havana had tracked the storm and read it correctly, including the westward turn into the Gulf. The US Weather Bureau, under a chief who regarded Cuban forecasting as a nuisance, had a standing policy restricting those cable warnings, and Washington did not move its own forecast until it was far too late to matter. Historians at Ohio State's Origins project lay out how the local forecaster, Isaac Cline, had publicly dismissed the notion that a hurricane could seriously damage Galveston.

Eight thousand people died of an institutional preference.

That's the real shape of the history, and it repeats with better technology each time:

• 1900 to 1960s: the knowledge exists in pockets and travels badly. Warning is a matter of who outranks whom.
• 1960s to 1990s: satellites and radar make the hazard visible from above, and sirens, radio and television push alerts out broadly but bluntly, to everyone or to no one.
• 2000s: the Common Alerting Protocol standardizes machine-readable alerts so any authority's warning can flow into any app or device. The National Weather Service still publishes its entire alert stream in that format, free and without an account.
• 2020s: models start producing forecasts for places that have no rain gauges, no river sensors and no national weather service worth the name. Coverage stops being a function of wealth.

That last step is the new one, and it's bigger than it sounds. Early warning used to be a thing countries bought. It's becoming a thing that exists. What's left over, once the forecast is free and the alert format is standardized, is plumbing: deciding which feed matters to you and what should happen when it changes, which is ordinary work for an agent with a trigger and a job.

The Early Warning Signal From a Flooded Ward in Kogi State

The clearest signal this year isn't a model benchmark. It's money arriving in the right place before the water did.

Kogi State, Nigeria sits where the Niger and Benue rivers meet, and the district of Ibaji floods so reliably that its communities become islands when the roads go under. In 2024 the nonprofit GiveDirectly wired AI river forecasts to a payment trigger instead of a press release. Working with hydrologists and local leaders, and watching Google's flood forecasts daily against what residents reported about water levels, GiveDirectly sent $105 each to more than 4,600 people, with the first 3,250 paid between August 29 and October 6. The flood peaked on October 21.

Weeks of warning, converted into cash, in the hands of people who knew exactly what to do with it. Recipients built raised platforms to keep food and seed above the water line, moved elderly relatives to higher ground, and stocked medicine. One of the photographs in GiveDirectly's own writeup shows a man named Felix with his household, sheltered, above the flood. Afterward, weekly income among recipients had roughly doubled, from about 4,500 naira to 10,000, and 93% said they felt better prepared for the next one.

Nobody in that story was rescued. They were warned, funded, and left to handle it themselves, which is a different and much better verb.

Who the Warning Still Misses

Now the part that keeps this column honest, because a forecast that only flatters the technology is worth nothing. The same Kogi program is also the clearest documented case of who falls through.

To move fast, the program enrolled people by text code rather than door to door, and verified them by phone. That saved an estimated $80,000 in field costs and about a year of time. It also meant the first filter on receiving aid was owning a working phone with signal, in a district where field teams found people climbing trees to get a bar. GiveDirectly says so plainly in its own writeup: phone ownership is lowest among the poorest, so the people the program most wanted to reach were the likeliest to be invisible to it. In August 2026, The New Humanitarian published an investigation into exactly that failure mode in Ibaji, documenting data mismatches and unreachable residents who watched the water take their farms with no transfer arriving.

Scale that up and you get the global picture. The World Meteorological Organization's Early Warnings for All initiative aims to cover every person on earth by the end of 2027. As of the latest global status reporting, 119 countries, roughly 60%, say they have a multi-hazard early warning system, which is a 113% increase since 2015 and also leaves about 40% of the world outside. The satellites see those places. The warning doesn't land there.

This is a familiar shape if you've been reading along. It's the same gap we hit with the models now watching the sky for methane leaks: detection raced ahead and the human machinery for acting on the detection stayed where it was. Sensing is cheap. Responding is political.

The Forecast for AI Disaster Prediction Through 2027

The honest read for the next 18 months: forecast quality and reach keep improving fast, post-disaster assessment collapses from weeks to hours, and the last mile improves slowly and unevenly. Two of three going your way, again.

Start with reach. Google Research reported in July 2026 that its flood forecasts now cover 2 billion people across more than 150 countries, with riverine forecasts out to seven days and urban flash flood nowcasts about a day ahead. On the storm side, the US National Hurricane Center used Google's WeatherNext model during the 2025 season, and Google DeepMind documented that it called Hurricane Melissa's historic Jamaican landfall five days out, which gave Jamaica's Met Service five days to tell people. For wildfire, the Earth Fire Alliance is building out the FireSat constellation specifically to catch small fires from orbit before they become the kind that make the news.

Then the aftermath, which is where the gains are almost absurd. A damage assessment workflow built with the UN Satellite Centre scored more than 385,000 buildings in Jamaica after Melissa, work that used to consume specialist weeks per activation. The UN's own July 2026 report, published through the ITU, is the clearest statement yet that this stops being a research story and becomes operational infrastructure.

By 2027 the question stops being whether anyone knew. It becomes why the person standing in the water didn't.

And the limits, plainly. Earthquakes remain largely unforecastable; what works there is detecting the first shaking and racing the wave outward, which buys seconds, not days. Models trained on global data can still miss local behavior, and the pilots that fed in local sensor readings improved sharply, which tells you the global model alone isn't enough. Flood coverage is widest for rivers, thinner for flash floods, thinnest for compound events where fire scars meet rain. None of this replaces a siren, a neighbor, or a plan. And the structural thing worth naming: this is still climate adaptation, which means it's the mop, not the leak. Better warnings are not a substitute for cutting emissions, and anybody selling them that way is doing a magic trick. The forecast infrastructure is also, mercifully, mostly dull recurring reporting work wearing a dramatic hat, which is why an individual can plug into it at all.

The Countermove: Build a Local Hazard Watcher

The street finds its own uses for things. Every agency and lab above is pushing warnings out through open, free, machine-readable feeds, and almost nobody is reading them. That asymmetry is yours to exploit.

Do the boring, load-bearing things first. None of these need a model, and they outperform anything you build:

• Turn on the alerts your phone already has: Wireless Emergency Alerts, earthquake alerts if your region has them, and your county's opt-in notification list, which is usually a webpage from 2011 and works fine.
• Learn your actual hazards: look up your address on the flood map, the wildfire risk map, and your state's hazard mitigation plan. Most people are exposed to two or three things, not twenty.
• Write the plan on paper: where you'd go, which route if the obvious one floods, who you'd call, and what you'd grab. Paper survives a dead battery.
• Keep one thing charged: a battery bank and a radio. The fanciest forecast in the world loses to a phone at 3%.

Then build the watcher. It's a scheduled agent pointed at your coordinates and at public feeds only: the National Weather Service alert API for watches, warnings and advisories, Google Flood Hub for river stage, your state or provincial fire agency's perimeter and incident feeds, an air quality endpoint for smoke, the USGS earthquake feed if you're somewhere that shakes, and your utility's outage map. It pulls on a schedule, compares against the last run, and discards everything unchanged, which is the same move as collapsing six dashboards into one short report.

Suppression is the entire product. The watcher stays silent on a normal Tuesday. It speaks when a watch becomes a warning, when a fire perimeter moves inside a distance you chose, when a river crosses the stage that matters for your street, when the smoke number crosses the line where you close the windows. One message, with the number, the threshold it crossed, and the one action you already decided on. Everything else, it eats.

One rule, and it isn't optional. Point it at hazards, never at people. Conditions, infrastructure, weather, water, smoke, shaking. Not who evacuated, not who stayed, not who was where. A tool that watches the sky protects a neighborhood. A tool that watches residents during an emergency is a cage that arrived in a helpful font, and this column does not build those. The agent workflow directory is full of monitors aimed at systems for exactly this reason.

The Horizon for Early Warning That Reaches Everyone

Here's the future worth building, with the priorities in the right order. A better flood alert does not outrank ending the wars burning right now, stopping the atrocities we can all see, or pulling the climate back from the line it's crossing. It's a piece of that third fight, and an honest one: the warming is already here, the storms are already stronger, and adaptation is what you owe the people living through the gap between now and the repair.

So picture the ordinary version, which is the only version that counts. It's a Thursday in 2029 in a town with no weather service of its own. A phone buzzes in a market stall at mid-morning: the river crests Saturday, here's how high, here's which roads go first. Nobody paid for that forecast. It came from a model that learned the basin from orbit because somebody decided ungauged places deserved forecasts too. By Saturday the seed is up on platforms, the goats are on the ridge, the grandmother is at her nephew's house on the hill, and the money to do all of that arrived on Tuesday because the same forecast tripped the same trigger that paid Felix's household in Kogi.

The water comes. The town gets wet. Nobody dies.

Notice what isn't in that picture. No command center deciding who gets to leave. No registry of who went where. No roadblocks, no permission, no one's movements logged in the name of keeping them safe, because the tools were only ever pointed at the river. The sovereignty runs the same direction as a block that can still hear itself with the towers down: the warning belongs to the person in the path of the thing, and so does the decision about what to do with it.

A hundred and twenty-six years ago, the people who knew a hurricane was coming for Galveston were in Havana, and the men who could have passed on the message decided they'd rather not. Every honest version of this future is just that message getting through. We finally built the wire. Now we have to use it on purpose.

Before Monday: Your Local Hazard Readiness Checklist

You've got a weekend. None of this needs all of it, and the first three take about twenty minutes together.

1. Look up your own address on your national or state flood map and wildfire risk map. Write down the two or three hazards you're genuinely exposed to and ignore the rest.
2. Check that Wireless Emergency Alerts are on in your phone's notification settings, and sign up for your county or municipal alert list while you're there.
3. Open the National Weather Service alert feed for your area once, by hand, so you've seen the raw thing the agent will be reading.
4. Decide your thresholds before you automate anything: how close a fire perimeter is too close, which air quality number closes the windows, which river stage means you move the car.
5. Build the watcher with the prompt above, pointed only at hazard and infrastructure feeds. Run it for a week and delete every source whose messages you skimmed past.
6. Write the one-page paper plan: route out, backup route, two phone numbers, where the documents are, who gets checked on first.
7. Tell one neighbor what you built and what your thresholds are. A warning one person receives is a hobby. A warning a street receives is a system.

Frequently Asked Questions

• How do AI disaster early warning systems work?+
  They learn the relationship between conditions and outcomes from decades of satellite and sensor data, then run that learned relationship forward. Google's river flood models forecast up to seven days ahead across more than 150 countries, weather models now beat older physics-only forecasts on cyclone tracks, and satellite constellations like FireSat detect small fires from orbit. The models produce the forecast. Governments, agencies and apps still have to carry it the last mile to the person standing in the water.
• Can I get AI flood and fire warnings for my own address for free?+
  Mostly yes, in the United States and in a growing list of other countries. The National Weather Service publishes every watch, warning and advisory through a free public API that needs no key and no account. Google Flood Hub shows river and flash flood forecasts for much of the world at no cost. Air quality, earthquake and wildfire feeds are similarly open. The gap is not price. The gap is that nobody reads six dashboards every morning, which is precisely the kind of job a scheduled agent exists to take off your hands.
• What is a local hazard watcher agent?+
  It's a scheduled agent that checks the official hazard feeds for your coordinates, compares what it finds against yesterday, drops everything unchanged, and messages you only when something new or escalating appears. Watches become warnings. A fire perimeter moves closer. A river crosses a stage threshold. The point is suppression, not summary: a watcher that pings you every day gets muted within a week, and a muted watcher is worse than none at all. If you want it to behave predictably, it helps to write the thresholds down as a spec before you build.
• Does AI early warning really save lives, or is it just better forecasting?+
  The evidence is strongest when a forecast is wired to an action rather than to a notification. GiveDirectly sent cash to more than 4,600 people in Kogi State, Nigeria weeks before the 2024 flood peak, triggered by AI forecasts, and recipients used it to evacuate and protect what they owned. Where no action is attached, a better forecast mostly produces a better-informed casualty. The forecast is the cheap part now. The plan attached to it is the part that still has to be built.