AI & Climate

Models can help interpret complex data, but their benefits and costs need to be measured together. A prediction should state its location, time horizon, uncertainty, and intended decision.

Key takeaways

• Specify place, horizon, and decision.
• Test distribution shifts and rare events.
• Account for both computation and downstream impact.

Define the physical or policy outcome first. Forecasting a local hazard, optimizing building energy, and classifying satellite imagery have different data and error requirements. Use time-aware evaluation when the future is the target, and test unusual events rather than only average conditions. Check measurement quality and geographic coverage. A model trained in one climate or sensor configuration may not transfer to another. Missing observations and changes in instrumentation can create apparent trends. Report uncertainty and the consequences of missed or false alerts for the communities using the information. Measure resource use as part of the system. Training and serving consume energy, while an optimized workflow may reduce energy elsewhere. State the boundary and assumptions of any comparison; a model’s compute estimate is not automatically a full lifecycle assessment. Keep decision authority clear for emergency, infrastructure, and environmental actions. Preserve source observations and communicate when a forecast is outside the evaluated range.

1. Imagine a model trained on ordinary weather days and evaluated only on average rainfall.
2. It performs well on routine days but misses the rare storms that matter most to emergency planners.
3. Add representative extremes, report uncertainty, and define a safe escalation path before using the forecast.

The hypothetical example shows why average error can hide climate-relevant failures.