SWITCHED ON
The daily technology series nobody asked for but everyone needed
The Early Warning: AI, Disaster Prediction, and Climate Adaptation
AI weather models can now outperform traditional forecasting on speed and, increasingly, on accuracy. Early warning systems save lives at a rate that is hard to overstate. The technology to predict disasters is improving faster than the technology to act on those predictions, and that gap is where people die.
A twenty-four-hour early warning before a cyclone or severe storm can cut resulting damage by thirty percent. This is one of the best-documented and most cost-effective interventions in the entire field of disaster management. The technology to deliver that warning is improving rapidly. Whether the warning reaches the people who need it, in time, in a form they can act on, is a separate and considerably less solved problem.
Yesterday we examined autonomous vehicles — the SAE levels that consumer marketing consistently obscures, Waymo's genuinely operational Level 4 service versus Tesla's still-Level-2 Full Self-Driving, the long-tail edge case problem that explains why the easy ninety percent of the engineering challenge arrived faster than the hard final ten percent, and the unresolved liability questions that will shape which technical approach succeeds commercially. Today we are turning to a domain where the stakes are measured directly in human lives saved or lost: disaster prediction and climate adaptation technology. AI weather forecasting, early warning systems, wildfire and flood prediction, and the uncomfortable truth that the gap between prediction capability and effective response capability is where preventable disasters keep happening.
01 — The AI Weather Forecasting Revolution
Weather forecasting has been transformed by AI models in a way that represents one of the most concrete and rapidly realised benefits of machine learning applied to a scientific domain. Traditional numerical weather prediction relies on solving complex physical equations describing atmospheric dynamics, requiring enormous supercomputing resources and taking hours to generate forecasts. Google DeepMind's GraphCast, released in 2023, demonstrated that an AI model trained on decades of historical weather data could generate ten-day global forecasts in under a minute on a single computer, with accuracy that matched or exceeded the traditional European Centre for Medium-Range Weather Forecasts model — widely regarded as the gold standard — on the majority of variables tested.
The speed advantage matters enormously for operational forecasting and for ensemble forecasting — running many slightly varied simulations to assess the range of possible outcomes and the confidence level of a prediction. What took a traditional supercomputer hours, an AI model can do in minutes, enabling forecasters to run far more ensemble members and produce probabilistic forecasts with better-characterised uncertainty. NVIDIA's FourCastNet, Huawei's Pangu-Weather, and several other AI weather models have demonstrated similar capabilities, and major meteorological agencies including the UK Met Office and the US National Weather Service have begun integrating AI models into their operational forecasting pipelines, typically as a complement to rather than full replacement for traditional physics-based models.
AI weather models learn statistical patterns from historical data rather than simulating the underlying physics. They are extraordinarily good at predicting weather that resembles patterns in their training data and meaningfully less reliable for genuinely unprecedented events — which is precisely the category of extreme weather that climate change is making more common. This is not a flaw to dismiss. It is the central limitation that operational meteorologists are actively working around.
02 — Early Warning Systems That Actually Work
The evidence base for early warning systems as a cost-effective disaster intervention is among the strongest in the entire field of disaster risk reduction. The UN's Early Warnings for All initiative, launched in 2022 with the goal of ensuring every person on Earth is covered by an early warning system by 2027, was motivated by World Meteorological Organization data showing that while disasters have increased fivefold over the past fifty years, the associated death toll has fallen threefold over the same period — a reduction attributed substantially to improved early warning and forecasting capability, even as the underlying hazards have intensified.
Bangladesh's cyclone early warning system is among the most extensively documented success stories in disaster risk reduction. Following Cyclone Bhola in 1970, which killed an estimated 300,000 to 500,000 people, Bangladesh invested heavily in a community-based early warning system combining satellite and radar forecasting with a network of volunteer-operated sirens and megaphones reaching coastal communities, alongside an extensive network of cyclone shelters. Cyclone Amphan in 2020, a storm of comparable intensity to Bhola, caused fewer than 100 deaths in Bangladesh — a reduction of several orders of magnitude attributable substantially to the early warning and evacuation infrastructure built over five decades.
The technology underlying modern early warning systems has advanced significantly: satellite-based flood prediction using AI analysis of rainfall and terrain data can now provide flood warnings days in advance with street-level precision in some deployments. Google's Flood Hub, using AI models trained on hydrological data, provides flood forecasting for river basins across more than 80 countries, including many that previously had no flood forecasting capability at all. Wildfire prediction systems combining satellite imagery, weather data, and AI models analysing vegetation dryness and fire spread patterns are being deployed by fire services in California, Australia, and increasingly in the Mediterranean region, where wildfire seasons have lengthened and intensified significantly due to climate change.
03 — The Last Mile Problem
The most sophisticated prediction technology delivers no benefit if the warning does not reach the people at risk, in a form they can understand and act on, with enough time and enough means to actually evacuate or take protective action. This "last mile" problem is where the gains from improved prediction technology are most likely to be lost, and it is a problem of infrastructure, communication, and social context rather than prediction accuracy.
The 2023 wildfires in Maui, Hawaii, which killed at least 100 people, exposed catastrophic last-mile failures despite reasonably accurate fire weather prediction. The county's outdoor warning siren system — designed primarily for tsunami warnings — was not activated during the fire, based on a judgment that sirens might cause confusion or be associated with tsunami protocols that would send people toward the fire rather than away from it. Emergency alerts sent via mobile phone were undermined by widespread power and cellular network outages caused by the fire itself. Many residents received no effective warning until the fire was already upon them. The prediction systems had identified extreme fire weather risk in advance. The warning delivery and evacuation infrastructure failed catastrophically.
This pattern — prediction outpacing response infrastructure — recurs globally and disproportionately affects lower-income and rural communities. Early warning systems depend on functioning communication infrastructure (which fails in many disasters precisely when it is most needed), on communities having received adequate preparedness training to know how to respond to warnings, on having accessible evacuation routes and shelter capacity, and on warnings being delivered in languages and formats that all affected populations, including those with disabilities or limited literacy, can act upon. Closing this gap requires sustained investment in physical infrastructure and community preparedness that is less technologically glamorous than AI prediction models and considerably more important for actual outcomes.
04 — Climate Adaptation Infrastructure Beyond Prediction
Prediction and warning are only the first stage of climate adaptation. The infrastructure investments required to actually reduce vulnerability to climate hazards — sea walls and managed retreat for coastal flooding, building codes and urban design for extreme heat, water infrastructure for drought resilience, and fire-resistant building materials and defensible space requirements for wildfire-prone areas — require capital investment at a scale that dwarfs the cost of prediction technology and that most vulnerable communities, particularly in low-income countries, cannot afford without significant external support.
The climate adaptation finance gap is stark: the UN Environment Programme estimates that adaptation costs in developing countries could reach $215 to $387 billion annually by 2030, while actual international adaptation finance flows have been a small fraction of this figure. The Loss and Damage fund established at COP27 in 2022, intended to provide financial assistance to vulnerable countries for climate impacts they did not cause and cannot afford to address, has been pledged at levels far below the estimated need, and disbursement has been slow. The technology to predict and even to design resilient infrastructure exists. The financing to actually build it at the scale required does not, for the populations who need it most.
05 — What Adequate Climate Adaptation Would Actually Require
The honest assessment of climate adaptation technology, several years into intensifying climate impacts, is that prediction and early warning capability has improved dramatically and genuinely saved lives, while the harder, more expensive, and less technologically glamorous work of building resilient physical infrastructure and ensuring warnings translate into effective action remains severely underfunded relative to need, particularly in the lower-income countries that face the most severe climate impacts despite having contributed least to causing them.
This is not an argument against continued investment in AI weather prediction and early warning technology — that investment has a strong evidence base and should continue. It is an argument that prediction technology without adequate investment in response infrastructure and capacity produces a specific and predictable failure mode: increasingly accurate warnings that arrive in communities without the infrastructure, resources, or capacity to act on them effectively. The Maui fires demonstrated this failure mode in a wealthy country with substantial resources. The gap is considerably wider in the low-income countries facing the most severe and most frequent climate disasters, where the technology gap that compounds the resource gap remains one of the most significant and least adequately addressed aspects of the global climate response.
Tomorrow we are going somewhere considerably more contested — gene drives and the technology to deliberately and permanently alter wild populations of species, starting with the campaign to eliminate malaria-carrying mosquitoes. The science is real. The ecological and ethical stakes of permanently altering an entire species are not small. See you then.
Switched On is a daily technology series covering the ideas, systems, and arguments shaping the digital world. Opinionated. Witty. Occasionally wrong. Always worth the argument.
