This week, I want to make an argument that runs against the current of almost everything being written about AI and climate change.

The breakthroughs are real. GraphCast and Aurora are remarkable. AI is going to reshape how we model planetary systems — that much seems certain.

But certainty, it turns out, is exactly the problem.

Climate science spent decades learning to hold complexity — to produce ensembles of possible futures rather than a single confident path, to flag uncertainty rather than smooth it away, to mirror the nonlinear, feedback-driven reality of the systems it studies.

Most AI is built to do the opposite. And if AI becomes the primary interface between climate models and the people making decisions, something important gets lost in that translation.

In this piece, I work through what ecological intelligence actually looks like, why pattern recognition breaks at thresholds, and what a genuinely different approach to climate AI might be designed to do — using a restoration project in China as a model for what adaptive, feedback-driven intelligence looks like in practice.