What Climate Models Got Wrong and What AI Might Fix

In 1990, the IPCC’s First Assessment Report projected that global mean temperature would rise by approximately 0.3°C per decade under a “business as usual” emissions scenario. The actual observed warming over the subsequent three decades came in at around 0.2°C per decade, meaningfully lower than the projection. Climate skeptics used this gap to argue that the models were running hot and that the whole enterprise was unreliable.

They were wrong about what the discrepancy meant. The models were projecting one scenario; the world followed a different emissions pathway, partly because the ozone treaty reduced some potent greenhouse gases and partly because industrial transitions in China and elsewhere created temporary aerosol cooling. When researchers went back and ran the 1990 models on actual emissions, the predictions were quite close to observations.

This episode is worth revisiting because the coming wave of AI climate modeling will generate similar misinterpretations. AI models will be faster, they will produce more visually striking output, and they will be wrong in different ways than physics-based models, ways that are less intuitive and harder to detect. Understanding the improvement potential requires first understanding the failure modes of what came before.

General Circulation Models (GCMs), the physics-based computational engines that have powered climate science for sixty years, work by dividing the atmosphere, ocean, land surface, and ice into three-dimensional grid cells and solving the equations of fluid dynamics, thermodynamics, and radiative transfer at each cell boundary. The most advanced current generation—CMIP6 models—typically run at horizontal resolutions of 50-100 kilometers. That means a single grid cell covers an area larger than a medium-sized city.

At that resolution, you cannot directly simulate individual clouds. Clouds—particularly the low marine stratocumulus sheets that cover roughly 20% of the ocean surface—are the dominant source of uncertainty in climate sensitivity estimates. Whether these clouds thin, thicken, or shift in response to warming is a first-order question for how much warming a given amount of CO2 produces. The models handle clouds through parameterization: statistical approximations derived from observations and theory that stand in for explicit simulation.

Different research teams make different parameterization choices. The spread in climate sensitivity across CMIP6 models—roughly 2.5°C to 5.5°C warming per doubling of CO2—reflects primarily differences in cloud parameterizations. That’s an enormous range. The top of that range represents a materially different world than the bottom.

This is the specific problem that higher-resolution AI modeling is attacking. Google DeepMind’s GraphCast, released in 2023, and ECMWF’s AIFS (Artificial Intelligence/Integrated Forecasting System) are trained on reanalysis datasets—reconstructed historical observations blended with model output—to predict atmospheric state directly. GraphCast predicted the track and intensity of Hurricane Lee’s September 2023 landfall more accurately than the conventional ECMWF model ten days in advance.

But GraphCast and its immediate successors are primarily weather models, not climate models. The distinction matters. Weather prediction needs to be accurate for a specific location at a specific time over a 10-day horizon. Climate projection asks what the statistical distribution of weather will look like over decades under different emissions scenarios. Those are fundamentally different tasks, and skills that transfer well between them are not guaranteed.

The initiative with the most direct ambition to close this gap is NVIDIA’s Earth-2 project, announced in 2021 and in active development since. The goal is to build an AI-powered simulation of Earth’s climate at kilometer-scale resolution—fine enough to explicitly resolve convection, which means fine enough to simulate clouds directly rather than parameterize them. Running that at global scale with sufficient temporal coverage to sample climate statistics would require, in the original estimates, roughly 3 million A100 GPUs running for several years.

That’s an extraordinary statement about the computational cost of reducing cloud uncertainty. It also neatly previews the AI-environmental paradox: the same models being developed to improve our understanding of climate change require massive carbon-intensive compute infrastructure to build and run. NVIDIA acknowledged this tension without resolving it. The company’s position is essentially that the scientific value of better climate models justifies the near-term carbon cost. That argument may be correct; it’s not obviously so.

The more tractable near-term application is using AI as a downscaling engine on top of existing GCMs. The approach works like this: run a coarse-resolution physics model at global scale to produce the large-scale atmospheric circulation patterns, then use a trained neural network to downscale the output to high resolution for a specific region. The AI learns the relationship between coarse-scale patterns and fine-scale local outcomes from historical observations.

The ClimateGAN approach developed by researchers at Mila (the Montreal AI institute) and collaborators did exactly this for precipitation. The models learned to generate statistically realistic high-resolution rainfall fields from coarse GCM output, producing downscaled projections that matched local observational records better than traditional statistical downscaling methods.

This matters enormously for adaptation planning. A national-scale projection that tells you precipitation will decrease by 10% in the Mediterranean region over the next 50 years is nearly useless for deciding whether to expand irrigation infrastructure in a specific river basin in southern Spain. A downscaled projection that provides plausible distributional estimates for that specific catchment is actionable.

Extreme weather prediction is where the immediate economic stakes are highest. The global cost of weather and climate disasters in 2025 exceeded $600 billion according to Swiss Re estimates—and that number reflects only insured losses. The uninsured losses in lower-income countries are larger.

The limiting factor for extreme event prediction is not the average behavior of the atmosphere but its tails. A model that perfectly predicts average temperature and precipitation while underestimating the probability and intensity of heat domes, atmospheric rivers, and compound drought-heat events is only marginally useful for risk management. The extremes are where the damage accumulates.

AI approaches have shown genuine progress here. DeepMind’s ensemble storm-track model, tested against 30 years of Atlantic hurricane data, identified structural predictors of rapid intensification—the dangerous process where a hurricane jumps a category or two in 24 hours—that physics-based models had systematically missed. The predictors turned out to involve sea surface temperature gradients at spatial scales smaller than the GCM grid resolution. A model that can’t see those gradients can’t use them as predictors.

This is precisely the kind of pattern AI is good at finding: nonlinear relationships across multiple spatial scales in high-dimensional data. The danger is that it will also find spurious patterns—statistical artifacts that worked in the training period but fail when the climate shifts into states not well-represented in historical data. Extreme events, almost by definition, are underrepresented in historical records. A model trained to predict them is working at the edges of its training distribution.

The honest assessment from the researchers building these systems is that AI climate modeling in 2026 is at the stage where it is clearly useful for specific applications—extreme event prediction, downscaling, emulating expensive physics models for scenario analysis—but has not yet demonstrated the reliability needed to serve as a replacement for physics-based models in long-term projections.

There is a deeper problem that neither physics-based nor AI climate models can fully solve: the climate system is being pushed into states with no historical analog. Paleoclimate records—ice cores, sediment layers, tree rings—provide data from periods when CO2 was at current or higher concentrations, but those were geological timescales with different continental configurations, different biomes, and different starting conditions.

The unprecedented rate of change is itself a source of model uncertainty. The climate system has absorbing feedbacks and tipping points—Arctic sea ice, Amazon dieback, permafrost methane release—that are represented poorly in current models and may be represented even more poorly in AI models trained primarily on the historical record. When the system crosses a tipping point, the statistical relationships the AI learned from pre-tipping data may be not just inaccurate but fundamentally inapplicable.

This is not an argument against AI climate modeling. It’s an argument for epistemic humility about what the models can and cannot tell us.

The appropriate use of AI in climate science is as a set of powerful tools in a scientific process that remains grounded in physical understanding: faster emulation of expensive physics runs, better extraction of signal from observational data, more skillful downscaling, improved extreme event prediction within historical distributional ranges. That is a large and valuable contribution.

What it is not is a shortcut past the hard problem of understanding a complex system that is changing faster than human institutions are designed to track. The 1990 IPCC models were criticized unfairly for getting the right answer for the wrong reasons. The AI models of the 2020s will face a more serious version of the same challenge: getting impressive-looking answers that prove unreliable exactly when reliability matters most.

The atmosphere doesn’t run on the training distribution. It runs on physics.