AI is moving into climate decisions: who gets an early warning, which villages receive anticipatory cash transfers, where adaptation money goes, which farms are flagged for deforestation. When those systems are wrong, the cost falls on people, often people with the least power to challenge it.
This is the side of AI I want to spend the next decade on.
Where AI is entering climate policy
- Early warning: forecasting floods, heat, and landslides (flood systems).
- Targeting: deciding who receives support before or after a disaster.
- Compliance: flagging land or supply chains as high-risk (EUDR).
- Planning: prioritising infrastructure and adaptation investment.
- Monitoring: emissions, land use, and environmental violations.
The risks
1. Data gaps become bias
Wealthier, urban, and well-studied areas have more data. Remote and poorer areas have less, so models are least accurate where vulnerability is highest. A flood model trained mostly on gauged rivers may under-warn ungauged communities.
2. False positives with real consequences
A deforestation flag on a smallholder's plot can cut them out of a supply chain. If there's no easy way to contest it, the model's error becomes their loss of income.
3. Opacity
If officials can't explain why a model prioritised one district over another, it's hard to hold anyone accountable.
4. Automation bias
People tend to trust model outputs too much, especially when they're presented on polished dashboards.
5. Vendor dependence
Governments relying on proprietary systems they can't inspect, maintain, or afford long-term.
Guardrails I'd insist on
| Guardrail | In practice |
|---|---|
| Human oversight | A named person accountable for decisions, with authority to override |
| Contestability | Affected people can challenge a flag or decision, with a clear process |
| Performance by group | Accuracy reported by region and population group, not just on average |
| Transparency | Public documentation of data sources, model purpose, and known limits |
| Data rights | Clear rules on what's collected, why, who can access it, for how long |
| Local capacity | Training and ownership so the system survives the vendor contract |
| Evaluation after events | Did warnings reach people? Who was missed? |
Existing frameworks
Useful references include the OECD AI Principles, UNESCO's Recommendation on the Ethics of AI, and, for EU contexts, the EU AI Act, which treats some public-sector uses as high-risk. I covered the developer view in EU AI Act for Developers.
A test for any climate AI project
Before deploying, ask:
- Who could be harmed if this is wrong?
- Would they know?
- Could they challenge it?
- Who is accountable?
If you can't answer all four, the system isn't ready.