The Productive Role of Error in AI-Driven Modeling
Sometimes the most useful thing an AI model does is fail.
- what
- We study how mistakes made by AI models in science can lead to new knowledge, starting with weather and climate.
- why
- AI is entering weather and climate science faster than anyone can explain how it works. Its failures show what a model has missed, and when it can be trusted.
- join
- Tell us about an AI model that went wrong, talk with us about your work, or come to our 2027 workshop.
What does a productive error look like?
Here’s a real one, from AI weather forecasting: how the model failed, what the failure revealed, and what science gained. Scroll to step through it.
forecast(ai_weather_model, days=365)
InstabilityError: forecast unstable beyond 10 days
FourCastNet, an AI weather model built on Fourier neural operators, a common design for AI weather models, made fast, accurate short-range forecasts. But beyond ten days, its forecasts fell apart.
diagnose(error)
It assumed the Earth was flat.
Researchers traced the problem to one component: a piece of math that treats the Earth like a flat map, which breaks down near the poles. Rebuilding that piece with math designed for a sphere gave a model that stayed stable for a full year of simulated weather.
Sources: Pathak et al. 2022, FourCastNet; Bonev et al. 2023, Spherical Fourier Neural Operators, ICML; O’Loughlin et al. 2025, Geoscientific Model Development, Sect. 4.2.
learn_from(error)
The error didn’t just get fixed. It left science better off:
- Understanding. Tracing the failure to one component showed how the model works, insight that explainability tools rarely provide.
- Antifragility: coming back stronger after a failure. The model went from ten stable days to a full year.
- Credibility. By diagnosing the failure in the open, researchers showed they understood a system that is hard to see inside.
Most discussions of AI in science ask how to eliminate errors. We ask what errors teach.
Sources: O’Loughlin 2023 and O’Loughlin et al. 2025 on understanding; Taleb 2012, Antifragile, on antifragility; Mayernik 2021 on credibility.
why_it_matters()
AI models are moving into everyday science. Since February 2025, the European Centre for Medium-Range Weather Forecasts (ECMWF) has used an AI model in daily forecasting, alongside its traditional model.
Climate is harder still. The future climate lies outside any data a model was trained on, so it can’t simply be checked against observations.
Knowing how these models fail, and what their failures reveal, is how scientists will know when to trust them.
Source: ECMWF 2025, “ECMWF’s AI forecasts become operational”.
run(project)
We’re talking with the people who build and use these models: scientists, software engineers and data specialists at the U.S. National Science Foundation’s National Center for Atmospheric Research (NSF NCAR), NASA’s Goddard Institute for Space Studies (NASA GISS) and Columbia University’s LEAP Center (Learning the Earth with Artificial Intelligence and Physics). We also read the reports, code and data records where errors get written down.
The questions
Error and understanding, antifragility (coming back stronger after a failure) and credibility in AI-driven modeling.
The workshop
Scientists, philosophers and scholars who study how science works (STS), working through real cases of error side by side.
The work
A framework for how errors produce knowledge, with articles for philosophers, STS scholars, scientists and the public.
The people: our team and advisory board.
Have an error story? We want to hear it.
Using AI in climate science, or in any other field? Tell us about a time a model went wrong and what it taught you. We also welcome collaborators.