By Ron Dembo
AI is making climate models faster. That is not in dispute.
The United Nations Secretary-General’s Scientific Advisory Board sets out what the technology now does well in its brief, AI and Climate Modeling. Emulation reproduces the behaviour of expensive physics-based models at a fraction of the cost. AI improves estimates of processes too small to simulate directly, such as small-scale convection. Downscaling turns coarse global output into local projections, and newer methods capture how hazards occur together: hot and dry conditions, or storms accompanied by heavy rainfall.
Some AI-based Earth system models now run stably for thousands of simulated years.
The most useful part of the brief is its caveat.
AI models perform best when the future resembles the data they learned from. The brief is direct about the consequence. As climate change pushes the Earth system into conditions with no historical precedent, models trained exclusively on historical climate data lose their footing.
Rare, high-impact events are the hardest to capture: extreme heatwaves, intense rainfall. They sit in the tails of statistical distributions.
That is the same failure that defines Climate 1.0. The architecture has changed. The calibration has not.
A model trained on the past, however fast, inherits the past’s range of outcomes. It will reproduce the centre of the distribution well. It will be least reliable in the tail, which is where the losses that break capital adequacy sit.
Speed does not fix this. Neither does resolution. A high-resolution projection of the wrong distribution is still the wrong distribution.
The brief adds a second point. Climate observations have major gaps across parts of the Global South, the deep ocean and the polar regions. AI systems inherit those gaps and can amplify them.
For a financial institution, this is not a scientific footnote. It is an exposure.
A model can return a flood depth or a heat index for an asset in a poorly observed region with the same number of decimal places as one in a well-instrumented city. The output looks equally precise. It is not equally grounded. The difference does not show up in the number. It shows up when the loss arrives.
That matters most where the financial consequence is largest: the credit quality of a loan book concentrated in under-observed regions, the insurability of assets where the historical record is thin, the capital held against exposures whose hazard estimates rest on sparse data. These are the places where a model trained on history has the least history to learn from.
The brief makes a third point that deserves more attention than it usually gets. A weather forecast can be checked against observations within days. A long-term climate projection cannot. We do not yet know the climate of 2050 or 2100, so there is no observed outcome to score a model against.
The brief’s conclusion is that public confidence in any climate model will depend on a well-established and recognized evaluation process, and that the pace of AI research makes this more urgent, not less.
For a CRO defending a climate number to a supervisor, this changes what counts as evidence. Back-testing against the past says little about a future with no precedent. A track record on next week’s weather says nothing about the tail in 2060.
What remains is the method. If the output cannot be validated against the future, the assumptions that produced it have to be inspectable: which climate pathways, which physical model, which hazard interactions, which damage functions. A model that cannot show its working is asking to be trusted on speed and resolution alone.
The brief does not conclude that AI is the wrong tool. It concludes that AI augments physics-based modelling rather than replacing it, and that hybrid approaches combining physical laws with machine learning are the most promising path. It also notes emerging work on AI models that include the statistical distribution of extreme events.
We agree. That is how our Climate Digital Twin (CDT) is built.
The CDT couples AI with a physics-based model and stochastic modelling of the climate future. The distinction that matters is what the AI learns from. We do not train it on history alone. We train it on a science-based stochastic future: 2,500 plausible climate futures, with compound and multi-hazard interactions captured at asset level.
The tail is in the training set because the training set was built to contain it.
The AI then does what the brief says it does well: speed, scale and resolution. The physics and the stochastic ensemble do what historical data cannot. They describe a climate that has not happened yet.
The choice of training set is not an academic detail. It determines what the model hands to the person making the decision.
A model trained on history produces a point estimate: an expected loss, a median projection, a single hazard score. It carries the confidence of a precise number and the blind spot of the record it came from. The tail is not underestimated. It is absent.
A model trained on a stochastic future produces a distribution. It shows the expected loss, and it shows how much worse the outcome can get, how often, and under which combination of hazards. That is the form the financial question takes. Capital adequacy, probability of default, loss given default and reinsurance attachment points are all set by the tail, not the centre.
A point estimate tells a board what is likely. A distribution tells it what it needs to be able to survive.
For a Chief Risk Officer (CRO), a Chief Actuary or a supervisor, whether a model uses AI is the wrong question. Most will, soon. The question is what future the model was trained on.
If the answer is the historical record, the output will describe the expected well and the tail poorly. If the answer is a physically grounded distribution of futures, the output can answer what a regulator will ask: how much, how often, under which conditions.
AI will make every climate model faster. The institutions that will be fine are the ones whose models were trained on the climate ahead, not the one behind.
Image credit: Berkeley Pit, Anaconda Copper Mind. Butte, Montana, 1985. © Edward Burtynsky, courtesy Nicholas Metivier Gallery, Toronto