How AI Can Help in Climate Change: What the Evidence Shows

Ask how AI can help in climate change and most answers arrive as a list: nine ways, five ways, ten ways. The lists are not wrong, but they put a satellite system verified by the United Nations next to a startup's own sales figure, in the same font, with the same confidence. A reader cannot tell which items have changed anything in the world and which are still a slide.

This guide sorts the uses by the strength of the evidence behind them. It is written for researchers, students, policymakers and the investors who fund them, and every figure links to the study or institution it comes from.

The short answer

AI helps with climate change in four ways. It makes emissions visible, it cuts emissions inside energy and transport systems, it forecasts hazards so people can prepare, and it informs where money goes. The reference survey of the field, Tackling Climate Change with Machine Learning, frames this as two jobs, reducing greenhouse gas emissions and helping society adapt, and stresses that the high-impact gaps are filled in collaboration with other fields rather than in machine learning alone.

What the lists leave out is that these four are at very different stages. Some have been verified against a control group or a physical outcome. Some run in daily operations but have not yet been measured for their effect on emissions or lives. Some exist only as estimates. The three tiers below follow that order.

Tier one: verified in the field

Finding methane leaks from space

Methane is the clearest win so far. The UN Environment Programme's Methane Alert and Response System (MARS) uses AI to scan data from over 30 satellite instruments, and its own report, Spotlighting Opportunity, says the models let its analysts process 12 to 15 times more data than they could without them. For large sources in arid regions, the models detect 78 percent of plumes at previously unseen sites with a false positive rate of 8 percent.

The outcome is what puts it in this tier. UNEP has confirmed more than 40 mitigation actions that MARS facilitated or verified, on sources estimated to have released 1.2 million tonnes of methane before they were fixed, comparable to the annual emissions of nearly 24 million petrol cars.

There is a catch, and it matters for everything that follows. Detection is not response. According to the World Economic Forum's summary of the same report, the response rate to these satellite alerts was just 12 percent in 2025, and the UN Secretary-General has called on governments to raise it to 80 percent by 2030.

Steering flights away from contrails

The condensation trails behind aircraft warm the planet, and AI can forecast where they will form. In a 2026 trial, contrail forecasts were built into American Airlines' flight planning software across 2,400 scheduled transatlantic flights, and the flights that followed the avoidance plans showed a 62 percent lower contrail formation rate than the control group. It is still a trial, not a fleet policy, but it was measured by satellite against controls, which is the standard the rest of the field should be held to. The full ledger of emissions cuts, including grids and buildings, is in our piece on AI in climate change mitigation.

Forecasts that reached farmers and floodplains

Adaptation has three deployments with measured outcomes. In 2025, India's agriculture ministry sent AI forecasts of the monsoon's onset to 38 million farmers with up to four weeks' notice, in a season when the monsoon stalled for weeks. An AI flood model described in Nature predicts extreme river floods in basins with no gauges at up to five days' lead time, and runs as a free public service in over 80 countries. In Senegal, voice advisories in local languages reached over 2,700 farmers, and a matched-village study found millet yields 41 percent higher among users.

What these have in common is not the model. It is a forecast, attached to a decision someone has to make, delivered through a channel they already use. We look at each in detail in AI in climate change adaptation.

Tier two: operational, impact not yet measured

Weather forecasting has changed faster than any other part of the field. On 25 February 2025 the European Centre for Medium-Range Weather Forecasts put its AI model AIFS into operations next to its physics-based system. It gains up to 20 percent on tropical cyclone tracks and uses roughly 1,000 times less energy per forecast. Google DeepMind's GraphCast, in its Science paper, beat the best operational deterministic system on 90 percent of 1,380 verification targets.

These systems are in daily use, so they belong above the estimates. They sit below tier one because better forecast scores are not yet the same as measured lives or losses saved, and because the models have a known weak spot. A 2026 study in Science Advances found that AI models consistently underestimated record-breaking heat, cold and wind, the very events early warnings exist for, and concluded they are not yet ready to be relied on alone for disaster management. Our article on AI in climate change prediction sets the wins and the failures side by side.

Power grids sit in the same tier. The International Energy Agency's Energy and AI report lists uses already in service at energy companies: AI fault detection that shortens outages by 30 to 50 percent, and sensors with AI management that could unlock up to 175 GW of transmission capacity without new lines.

Tier three: modelled and claimed

This is where the large numbers live. A 2025 peer-reviewed paper in npj Climate Action estimates that AI in power, food and mobility could cut emissions by 3.2 to 5.4 gigatonnes of CO2 equivalent a year by 2035. The same authors point out that earlier industry figures, from companies that sell AI, were not peer reviewed and did not disclose their full methods. The IEA's estimate is more cautious: broad adoption of existing AI could cut around 5 percent of energy-related emissions in 2035, which it calls far smaller than what is needed.

Below the modelled estimates sit the claims. The World Economic Forum's list of nine ways AI fights climate change includes an emissions platform said to help companies cut emissions by 20 to 30 percent. That figure comes from the vendor. Large providers report the same way: Google says five of its AI products enabled about 26 million tonnes of reductions in 2024, a self-reported number set against company emissions of 11.5 million tonnes. Neither is necessarily wrong. Both are claims, not measurements, and a funder can ask for the difference.

The debit side: AI's own footprint and the rebound

Any honest answer has to subtract. 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 in a high-growth case. That is below 1.5 percent of energy sector emissions, but it is one of the fastest-growing sources.

The larger risk is what AI is pointed at. A 2026 model in npj Climate Action treated AI as a productivity boost for every energy sector, oil, gas and coal included, and found that net emissions rise by 0.47 to 1.8 gigatonnes of CO2 a year. In that model, cuts only win when renewables gain four to five times more than fossil fuels.

The footprint that remains can be managed. A 2026 review in Energy Informatics finds that running AI workloads on low-carbon grids and at low-carbon hours cuts their emissions by 20 to 70 percent.

What actually limits the help

Three constraints come up again and again, and none of them is a better model.

1. Action after the alert. The methane system can see leaks; in 2025, most of its alerts went unanswered. The same pattern holds for weather warnings that never reach the farmer.
2. The future is not the training data. The UN Secretary-General's Scientific Advisory Board notes that AI models perform best when conditions resemble the past they learned from, and that a warming planet is moving into conditions with no precedent.
3. Money. UNEP's Adaptation Gap Report 2025 estimates adaptation finance needs in developing countries at US$310 billion to US$365 billion a year by 2035, against US$26 billion of international public flows in 2023. The cheapest proven tools in tier one are cheap relative to that gap. How markets price the risk on the other side is covered in AI climate risk pricing.

Where the summit fits

Every tier-one result above needed the same three parties in one project: a research team, an operator who could act, and someone to pay for it. The airline, the ministry and the UN programme were as necessary as the model.

That is the room the EX Future Summit is built to hold. 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, with free online attendance for verified researchers and students. Its program includes an Impact track and a Global Financial Raise track on R&D funding, with hosted sessions matching universities to private partners, for projects that have a model and still need the operator and the capital.

FAQ

Can AI solve climate change?

No, and the strongest sources say so plainly. The IEA describes AI as not a silver bullet and estimates that even broad adoption would cut around 5 percent of energy-related emissions in 2035. AI makes specific tasks faster and cheaper; it does not replace policy, investment or emissions cuts in the physical economy.

What is the best proven example of AI helping the climate?

UNEP's methane alert system is the strongest so far. It has contributed to more than 40 verified mitigation actions on leaks that had released an estimated 1.2 million tonnes of methane. Contrail avoidance in airline trials is close behind.

Does using AI tools like chatbots harm the climate?

They add to electricity demand, and data centres are among the fastest-growing sources of energy sector emissions, though still below 1.5 percent of the total. Where and when the computing runs matters more than most users realise, since cleaner grids and timing can cut workload emissions by 20 to 70 percent.

How can students or researchers start working on AI for climate?

Start from a problem rather than a model. The Rolnick survey maps open problems by sector, and the most useful gaps are in data-scarce regions and in measuring outcomes, the step that moves a use from tier three to tier one.