Earlier we totted up AI's environmental bill — the energy, the water, the data centres. This is the other side of the ledger: where AI is genuinely helping the climate. Not the gauzy "AI for good" brochure version — the cases that survive a hard look, counted honestly, hype stripped out.

The flagship: AI now forecasts the weather better than the supercomputers

This is the strongest example, because the people AI beat have adopted it.

Google DeepMind's weather models, GraphCast and the newer ensemble model GenCast, can produce a 15-day global forecast in minutes on a single chip — work that takes a supercomputer hours. And they're not just faster: GenCast beat the European Centre for Medium-Range Weather Forecasts (ECMWF) — the gold standard — on 97% of the measures tested (Nature, 2024).

The clincher isn't the benchmark, it's the adoption: in 2025 ECMWF put its own AI forecasting model into live operation, running alongside its physics-based one — and says the AI version uses on the order of a thousand times less energy per forecast. Better forecasts mean better flood warnings, smarter renewable scheduling and earlier storm tracks. When the incumbent starts using the upstart, it's real.

(The honest footnote: these models are still trained on decades of physics-based weather data and need a traditional analysis to start each forecast. AI has overtaken physics on skill, speed and energy — but it's standing on physics' shoulders, not replacing it.)

The British one: halving solar forecast errors on the grid

Closer to home, the charity Open Climate Fix built machine-learning models that predict UK solar output far more accurately — halving forecast errors — and they're now in live production feeding National Grid ESO. Better solar forecasts mean the grid operator holds less fossil-fuelled backup "just in case." The reported saving: over £30 million and around 300,000 tonnes of CO₂ a year. Unglamorous, current, and concrete — the kind of win that actually moves the needle.

Catching the leaks you can't see

Methane is a fierce greenhouse gas, and a lot of it escapes from infrastructure nobody's watching. The French firm Kayrros trained machine learning on European Space Agency satellite imagery and found roughly 1,800 methane "ultra-emitters" worldwide — huge leaks largely missing from official inventories (Science, 2022). You can't fix what you can't see; AI made these visible.

It's also a fair place to be honest about limits: the much-hyped MethaneSAT, a dedicated methane-tracking satellite, lost power and was declared likely unrecoverable in mid-2025. The AI and the data live on; the spacecraft doesn't. Progress here is real but not frictionless.

The shorter list, fairly labelled

• Fusion: DeepMind and Switzerland's EPFL used AI to magnetically control the plasma inside a real tokamak reactor (Nature, 2022) — sculpting it into shapes human controllers struggle to hold. That's genuine, and genuinely cool. It is control and stability research, not "AI cracked fusion" — net fusion energy is still years and hardware away.
• Materials: DeepMind's GNoME predicted roughly 380,000 stable new crystal structures, some promising for batteries and solar. Note the word predicted — these are computational candidates, most not yet made or proven, and materials chemists have pushed back on how many are genuinely novel. A search tool, not a warehouse of new materials.
• Conservation: AI listens for chainsaws in rainforests (Rainforest Connection, 750,000+ hectares monitored), sorts millions of camera-trap photos thousands of times faster than people, and even predicts where Amazon deforestation will strike next (PrevisIA, ~73% accuracy).

The part the cheerleaders skip

Here's the catch that keeps this honest: the same technology behind every win above is also driving emissions up. The International Energy Agency expects data-centre electricity to more than double by 2030, to roughly the entire electricity demand of Japan, with AI the biggest driver — though that's still about 3% of global electricity, so don't catastrophise either. More pointedly, Google's emissions are up ~48% since 2019 and Microsoft's ~23% since 2020, both blamed on the AI build-out — even as both firms market AI as a climate saviour.

So the grown-up verdict isn't "AI will save the planet," and it isn't "AI is wrecking it." AI is a double-edged tool, and its net climate effect depends entirely on what we point it at. The forecasting, the leak-hunting, the grid models — those are the cases where it's genuinely earning its footprint. The honest work is making sure the helpful uses outrun the hungry ones.