AI in Climate Change Mitigation: Uses, Limits, Energy Cost

Most writing on AI in climate change mitigation falls into one of two camps. The first adds up the emissions AI could remove from power grids, buildings and transport, and arrives at very large numbers. The second adds up the electricity that AI data centres consume, and arrives at alarm. The two camps rarely cite each other, and so the reader is left to guess which one describes the real world.

This piece puts both columns on the same page, along with a third result that neither camp likes: a 2026 model in which AI raises net emissions. It is written for researchers, students, policymakers and the investors who fund them, and every figure links to the study it comes from.

What AI in climate change mitigation means

Mitigation is the part of climate action that reduces the greenhouse gases entering the atmosphere. It is distinct from adaptation, which helps people cope with the warming already locked in (we cover that in AI in climate change adaptation), and from forecasting, which we examine in AI in climate change prediction.

A 2025 paper in npj Climate Action groups AI's mitigation role into five areas: transforming complex systems such as grids, accelerating the discovery of technologies and materials, nudging behaviour, modelling climate interventions, and strengthening resilience. The first three do most of the measurable work, and they are where the evidence below sits.

Where AI is already cutting emissions

Estimates are cheap. Controlled trials and operating systems are not, so they come first.

Contrails: the clearest trial evidence

The condensation trails behind aircraft account for roughly 35% of aviation's warming impact, according to the 2022 IPCC figure quoted by Google Research. In 2023, American Airlines pilots flew 70 test flights using AI forecasts of the zones where contrails form, and satellite imagery showed 54% fewer contrails. The avoiding flights burned 2% more fuel, but because only a small share of flights needs adjusting, the fleet-wide fuel penalty could be as low as 0.3%, putting the cost at around $5 to $25 per tonne of CO2e.

The follow-up is what moved it from a stunt to an operation. In a 2026 trial, the forecasts were built into American's flight planning software across 2,400 scheduled transatlantic flights, and the flights that flew the avoidance plans showed a 62% lower contrail formation rate than the control group. It remains a trial, not a fleet policy, but it is the most rigorously verified mitigation result AI has produced.

Grids and renewables

The International Energy Agency's Energy and AI report lists the applications already in use by energy companies. AI improves forecasts of wind and solar output, which reduces curtailment. AI-based fault detection shortens grid outages by 30 to 50%. Remote sensors combined with AI management could unlock up to 175 GW of transmission capacity without building a single new line, which the IEA notes is more than the growth in data centre load it expects to 2030.

The npj Climate Action authors estimate that better grid management, raising the load factor of solar and wind by up to 20%, could cut power sector emissions by 1.8 GtCO2e a year by 2035. That makes the grid the single largest line on the credit side.

Buildings

A study in Nature Communications modelled US medium office buildings and found that AI, by lowering the cost of high-efficiency design, construction and control, could reduce their energy use and CO2 emissions by about 8% by 2050 under business as usual, and by about 19% when paired with efficiency policy. The point is less the number than the mechanism: AI mostly helps by making proven efficiency measures cheaper to deploy.

Transport routing

Google reports in its environmental paper on AI that fuel-efficient routing in Google Maps enabled more than 2.7 million tonnes of emissions reductions in 2024, and that five of its AI products together enabled about 26 million tonnes. Those are self-reported "enabled" figures from an AI provider, which is worth keeping in mind; the company's own total emissions in 2024 were 11.5 million tonnes.

The credit side: how big could the cuts be

Scale these deployments up and the estimates grow quickly. The npj Climate Action paper concludes that AI in just three sectors (power, food and mobility, together about half of global emissions) could reduce emissions by 3.2 to 5.4 GtCO2e a year by 2035 against business as usual. Mobility contributes 0.5 to 0.6 Gt of that, mostly through shared vehicles and cheaper batteries.

The same authors are candid about earlier figures. They note that the widely quoted industry estimates, 1.5 to 4% from Microsoft and PwC and 5 to 10% by 2030 from Google and BCG, were not peer reviewed, did not disclose their full methods, and were produced by companies that sell AI.

The IEA is more conservative. It estimates that broad adoption of existing AI solutions could cut emissions by around 5% of energy-related emissions in 2035. In its words, that is far larger than the emissions from data centres, and far smaller than what is needed to address climate change.

The debit side: AI's own energy cost

The IEA puts emissions from data centre electricity at 180 million tonnes today, rising to 300 million tonnes by 2035 in its base case and 500 million tonnes in its high-growth case. That stays below 1.5% of energy sector emissions, but data centres are among the fastest-growing sources. A typical AI-focused data centre uses as much electricity as 100,000 households, and the largest now under construction will use twenty times as much.

The local picture is sharper than the global one. Google's paper cites projections that data centres will use about 3% of global electricity by 2030, yet could consume 6.7 to 12% of US electricity by 2028, and the company's own data centre electricity consumption grew 27% in 2024. Averages hide the grids that feel the load first.

The rebound problem

Both sides of the ledger so far assume AI is pointed at clean uses. A 2026 study in npj Climate Action drops that assumption. Its authors model AI as a productivity boost for every energy sector, including oil, gas and coal, and find that net annual CO2 emissions rise by 0.47 to 1.8 gigatonnes, equal to 1.2 to 4.8% of 2024 energy-related emissions. In their model, enabled emissions exceed avoided ones whenever fossil producers gain at all, and net cuts require renewables to gain four to five times more than fossil fuels.

The IEA reaches a milder version of the same warning. Rebound effects, such as people leaving public transport for cheaper autonomous cars, could undercut the gains, and its report states plainly that AI "is not a silver bullet and does not remove the need for proactive policy."

Set side by side, the three results are not contradictory. They describe the same technology under different steering. AI applied to grids, buildings and flight paths cuts emissions; AI applied equally to extraction can cancel those cuts and more.

What decides the sign

If steering decides the outcome, the useful question is what the steering levers are. The evidence points to four.

1. Where the money goes. The rebound result is a statement about relative productivity gains. Funding that favours renewables, grids and efficiency over extraction is what moves the ratio past four or five to one.
2. Where and when AI runs. A 2026 systematic review in Energy Informatics finds that carbon-aware scheduling and workload placement can cut the emissions of AI workloads by 20 to 70%, and that the carbon intensity of the local grid often matters more than chip-level efficiency.
3. Measurement that survives scrutiny. The contrail work is persuasive because satellites verified it against a control group. Self-reported enabled emissions are not the same standard, and funders can ask for the former.
4. Policy that sets direction. Both the IEA and the rebound study conclude that without policy, market forces alone will not choose the low-carbon uses. The finance side of that question, how markets put a price on climate risk, is the subject of our piece on AI climate risk pricing, and the wider map sits in how AI can help in climate change.

Where the summit comes in

Every lever above sits between a research group and a funder. The contrail trials needed an airline, a research lab and a satellite dataset in the same project; the grid gains need utilities, universities and capital to agree on what gets measured.

That meeting point is what the EX Future Summit is built around. It runs 18 to 20 November 2026 online, in Las Palmas and in Bali, as a single thirty-hour broadcast between two islands twelve hours apart, and online attendance is free for verified researchers and students. Its program includes a Global Financial Raise track covering venture capital, national grants and Horizon Europe, and eight hosted sessions matching universities with private partners: working sessions rather than keynotes, for projects that need a research team and a capital partner at the same table.

FAQ

Does AI produce more emissions than it saves?

It depends on what AI is used for. Studies that assume AI is applied to power, food and mobility find the cuts outweigh data centre emissions by a wide margin. A 2026 model that also lets AI make fossil fuel production cheaper finds a net increase of 0.47 to 1.8 Gt of CO2 a year. The outcome is set by policy and funding, not by the technology alone.

Is contrail avoidance with AI worth the extra fuel?

The early evidence says yes. Adjusted flights burned about 2% more fuel, but because only a small share of flights needs adjusting, the fleet-wide penalty could be around 0.3%, for a cost of roughly $5 to $25 per tonne of CO2e avoided.

How can AI's own carbon footprint be reduced?

Mostly by choosing where and when it runs. Running workloads on low-carbon grids and at low-carbon hours can cut their emissions by 20 to 70%, often more than model-level optimisation achieves.

Is AI in climate change mitigation a good research paper topic?

Yes, because the literature is still split. A strong paper can test one of the open disagreements: industry estimates against peer-reviewed ones, enabled emissions against verified trials, or the rebound effect in a single sector.