By: Zannirah Rahman
The sky wrote its own algorithm this year. The rains arriving earlier than expected, rewrote every prediction that claimed to know them. By September, the Indus had once again broken its banks. The National Disaster Management Authority’s (NDMA) Monsoon SitRep reported an unsightly picture of swollen rivers, breached embankments, and flash floods rushing through Sindh and southern Punjab. By the NDMA’s own count, more than 160 people had died and 70,000 homes were damaged across 38 districts. Days later, the United Nations Office for the Coordination of Humanitarian Affairs’s Pakistan Flood Situation Update confirmed what satellites had already seen, nearly one-third of the affected villages lay outside “high-risk zones” previously mapped by predictive models. Those zones, it turns out, were blind spots.
In a country where nearly two-thirds of the yearly rainfall arrives through the South Asian monsoon, predictive accuracy in forecasting isn’t about showing off scientific skills, it’s a matter of survival. This explains institutional disparities within AI Climate modelling. As the era of AI for Climate prevails, Pakistan still finds itself at a dangerous crossroads: global algorithms promise precision, but they often misread the very sky they claim to understand.
Across the world, “AI for Climate” has become the buzzword of the decade. Governments promise machine-learning flood maps, automated early-warning dashboards, and data-driven drought forecasts. Silicon Valley calls it the age of digital resilience. But in Pakistan, where two-thirds of the year’s rainfall arrives in a chaotic, fragile monsoon, faith in the algorithm has started to look more like climate techno-solutionism, a belief that software alone can outsmart the storm.
Economist Xianfu Lu’s 2022 paper for the Asian Development Bank captures the irony perfectly. Despite billions pledged globally, most AI climate projects still thrive in rich countries, the ones least likely to face catastrophic floods. Meanwhile, developing nations like Pakistan are left with ‘data deserts’, thin, patchy networks of weather stations and missing archives. “The algorithms,” Lu writes, “work best where the rain matters least.” Intersecting this with the very crux of Ulrich Beck’s Risk Society, the idea of vulnerabilities being manufactured seems predominant. The hastening of production supports those in their safe havens, while those who least contribute to climate change are at the brunt of climate-induced disasters.
The imbalance isn’t just technical; it’s moral. When AI models are trained on Europe’s stable skies or Japan’s fine-grained weather logs, they struggle to make sense of the messy choreography that defines Pakistan’s monsoon: Himalayan snowmelt meeting Arabian Sea moisture, interrupted by wild western disturbances.
How can a machine trained on calm skies predict a storm it has never met?
Not all hope is lost. In a groundbreaking 2021 study published in the Asia-Pacific Journal of Atmospheric Sciences, Muhammad Adnan and colleagues built a homegrown forecast model using Pakistan’s own climate data. Their results were startling: with local wind and pressure inputs, their system predicted 75 percent of year-to-year rainfall changes in Pakistan’s north-western regions.
As the 2025 monsoon fades, Pakistan once again finds itself between two futures. One is driven by imported systems that misread its landscapes. The other is built on open data, regional cooperation, and local innovation. To reach the latter, Pakistan must invest in open climate databases, renewable-powered computing, and collaboration across South Asia’s river basins. Only then can the next forecast truly belong to the people under its clouds. Because until then, the algorithms will keep forecasting the sky, and Pakistan will keep drowning beneath it.
By contrast, imported systems like ERA5 or CCSM4, the global standards of climate science, often overestimated rainfall in winter and misplaced the intensity zones of the pre-monsoon. As Safdar et al. (2022) found, these foreign models “capture the physics but not the geography”, a poetic way of saying they miss where the water actually falls, and where it fails to.
During the 2022 floods, several of those global prediction networks raised alarms, but the deadliest deluge hit places their maps had labeled ‘moderate-risk’. Two years later, it happened again. This discovery remains pivotal in understanding the knowledge systems within the Global South, incorporating Mignolo’s “broader canon on thought”, within the western narratives of climate mitigation.
But sometimes, the best solutions come from those closest to the rain. In a small research lab in Lahore, Hammad et al. (2021) built an AI model trained on just three meteorological stations (Astore, Chillas, and Gilgit). The model, called a Wavelet-Coupled Neural Network, predicts daily rainfall using nothing more than past rainfall data. It wasn’t flashy. It wasn’t expensive. Yet it achieved over 90 percent accuracy, better than most foreign models built on supercomputers.
That success held a quiet message: Pakistan doesn’t need to borrow algorithms to understand its own sky. What it needs is investment in local data, its own sensors, coders, and open datasets. What experts call ‘data sovereignty’, the right to train, test, and trust your own numbers, isn’t a luxury anymore. It’s a matter of survival.
That question “who owns the sky” is now political. Pakistan’s climate data sits in silos. The Pakistan Meteorological Department guards its archives; NDMA releases updates but not the underlying numbers; private satellites collect images that rarely make their way back to local universities. A 2024 study in AI & Society found that over 70 percent of all climate–AI research originates in the Global North, often relying on datasets that exclude South Asia. That gap creates what scholars call ‘epistemic inequality’, a world where algorithms learn about climate change without ever learning from the people living through it. When the power to see the weather lies outside the countries most affected by it, even forecasts become a form of dependency.
Even when AI works, it comes at a cost few discuss: electricity. Cowls et al. (2021) warn that training large neural networks can emit as much carbon as the lifetime of five cars. For a country like Pakistan, where blackouts remain common, powering AI can mean choosing between running a model or running a hospital generator. AI, they argue, is a gambit, a risky trade of energy for insight. In Pakistan, that gamble can feel like asking the flood victim to pay for the forecast.
So, what’s the alternative? Not to abandon technology, but to make it human. As Lu (2022) explains, true adaptation blends digital systems with local wisdom using tools based on community mapping, indigenous forecasting, and on-ground observation. In the delta, fishermen have long known how to read the monsoon, through traditional epistemic knowledge systems. A certain wind direction, a heaviness in the air, the way sea foam gathers before a storm. Thus we can imagine an AI model that listens to that knowledge, fed not just by satellites, but by the people who have lived through the weather for generations. That would be real intelligence, one that doesn’t just predict, but understands.
As the 2025 monsoon fades, Pakistan once again finds itself between two futures. One is driven by imported systems that misread its landscapes. The other is built on open data, regional cooperation, and local innovation. To reach the latter, Pakistan must invest in open climate databases, renewable-powered computing, and collaboration across South Asia’s river basins. Only then can the next forecast truly belong to the people under its clouds. Because until then, the algorithms will keep forecasting the sky, and Pakistan will keep drowning beneath it.
The writer is a freelance columnist
Insightful Piece