When someone acknowledges that they don’t completely trust the model they just built, a certain kind of silence falls in the room. The quiet is louder than any conference keynote when I sit in those rooms, which are typically small academic seminars. Scientists studying climate change are familiar with their formulas.
They have dedicated their careers to statistical downscaling, Navier-Stokes, and partial differential equations. Then, in 2022 or so, the topic of discussion changed. ChatGPT took place. The term “transformative” began to be used by investors in inappropriate contexts. All of a sudden, those performing the meticulous, slow math started to experience something akin to professional pressure.
| Topic | Mathematical and AI-based climate modeling |
| Field | Climate science, statistics, machine learning |
| Key Agreement | 2016 Paris Agreement — limit warming to 1.5°C |
| Primary Body | Intergovernmental Panel on Climate Change (IPCC) |
| Notable AI Models | Pangu-Weather, GraphCast, AIFS, Earth-2, Aurora |
| Core Concern | Training data may not represent a future warmer climate |
| Reference Paper | Benestad (2026), posted on arXiv |
| Public Profile | Gresham College Professor of the Environment, Jacqueline McGlade |
| Open Question | Will algorithms complement, or quietly replace, mathematics? |
In the climate community, you can already feel it. “AI will not replace you, but a person using AI will” is a common phrase that has the rhythm of a threat disguised as guidance. In certain industries, that might be the case. However, it lands differently in the field of climate science. Weather forecasting tools like Pangu-Weather, GraphCast, and Aurora are the success stories that everyone cites. the climate. not the weather. The distinction seems academic until you realize that it is the root of the whole issue.
The atmosphere we’ve already experienced is used to train weather models. Almost by definition, climate change is the atmosphere we haven’t experienced yet. The patterns of a world that is now vanishing are learned by an algorithm fed decades’ worth of historical temperatures. The more meticulous researchers feel that this is the silent weakness that no one wants to include on a slide show.
In a recent paper that was uploaded to arXiv, Rasmus Benestad used a metaphor that really resonated with me: a cuckoo laying its eggs in another bird’s nest. Neural networks that no one can thoroughly examine are gradually replacing decades of statistical downscaling work that has been tested and refined. Although the black box problem is not new, it assumes a different significance when applied to a planetary system. It’s amusing when Google’s AI summaries dream up a recipe. It’s policy when a climate model predicts a monsoon.
The uncomfortable reality of the carbon footprint is another. Large model training necessitates hot data centers that rely on fossil fuel-powered grids for power. Predicting climate warming using climate-warming infrastructure has the feel of a joke that no one is saying aloud.
However, the mathematics community is also not without fault, and this is where being honest is important. The IPCC reports are based on massive computer models that are assembled from physics, statistics, and assumptions that don’t always hold up to close examination. The Paris 1.5-degree target is surrounded by a cloud of uncertainty that most decision-makers would rather avoid talking about. Jacqueline McGlade, a professor of the environment, was appointed by Gresham College in part because the public should be aware of how precarious some of this ground actually is.
The truth is that traditional mathematics and artificial intelligence ought to collaborate. Their strengths are different. Uncertainty is handled neatly by statistics. The laws that the data must follow are enforced by physics. Patterns that humans miss are discovered by machine learning. The accountant’s reasoning that one can replace the other in order to save money is the threat, not AI per se.
It’s difficult not to consider how frequently we’ve confused speed for comprehension as we watch this play out. Climate change won’t wait for us to determine the best course of action. However, the algorithm that trends the loudest this quarter won’t solve it either. We still have a planet to figure out somewhere between the neural networks and the equations.
