How could AI amplify the climate crisis?

Will AI push global emissions up or bring them down? For 3 years, the question has pitted the terawatt-hours of data centres against promises of optimisation. In February 2026, 11 researchers, from Imperial College, Oxford and the Potsdam Institute among others, published online a study that quantifies the 2 together↗.

This 11% figure calls for 3 caveats. It covers cumulative CO2 alone, from 2025 to 2040, and should be read as a difference from a reference scenario in which AI has no effect: it says neither what will happen nor what share of emissions in 2040 comes from AI. It is the least favourable of 4 scenarios, built from the estimates of 17 experts. Finally, the study is a preprint that has not yet been peer-reviewed, and its authors ask that it be treated as exploratory work.

Its value lies in adding together what AI consumes and what it leads others to do. The authors suggest that the indirect effects could outweigh electricity, which receives almost all the attention.

The scenarios differ mainly in the rules that govern AI. In 3 of the 4, AI is powerful and ubiquitous. What varies is AI governance, the room given to climate policy and the source of the electricity.

In the least favourable scenario, emissions in 2030 are 8% higher than in the reference scenario, and then fall by only 4% over 10 years, while the reference falls by 12%. According to the authors, these results suggest that, within the range explored, the risks of an increase outweigh the chances of a decrease.

AI's electricity demand is the most studied effect. The experts surveyed put it at 1,395 TWh in 2030 and 3,224 TWh in 2040, as median values. That is 4.5% and then 10.4% of world electricity consumption in 2024.

The climate effect then depends on where that electricity comes from. If climate policies are not extended beyond 2030, high demand supplied by dedicated fossil-fuel plants adds 6% to cumulative emissions. Drawn from the grid at least cost, it adds 2.7%, because the cheapest new capacity is renewable. Supplied by nuclear or renewables, it adds nothing in the model.

According to the Shift Project↗, more than half of the electricity consumed by data centres worldwide came from fossil sources in 2024, and their footprint could reach 920 million tonnes of CO2e per year in 2030.

The climate scientist Valérie Masson-Delmotte, who co-chaired Working Group I of the IPCC from 2015 to 2023, makes the same observation in an overview of data centres↗ presented on 30 May 2026. She sees it as a challenge for using low-carbon electricity in place of fossil fuels, points to a massive revival of gas infrastructure in the United States, and notes that Google's and Microsoft's decarbonisation targets for 2030 are compromised by generative AI.

The origin of the electricity therefore matters as much as its quantity. That is why the hub publishes the carbon intensity of electricity country by country, and why the same question put to an AI does not carry the same weight depending on the grid that powers it.

An AI that helps find a deposit, drill faster or operate a well better makes oil and gas cheaper to produce, and a fall in costs can push production up. This mechanism, flagged as early as 2024 in a commentary in the journal Nature↗, was quantified in August 2026 by a study published in npj Climate Action↗.

The authors feed the productivity gains that AI would bring to the energy sectors, fossil and renewable, into a model of the global economy. The emissions enabled by gains in the fossil sector, before deducting avoided emissions, range from 0.6 to 2.4 billion tonnes of CO2 per year. That is 3.3 to 13.3 times the International Energy Agency's estimate for data centres, 0.18 billion tonnes, which the authors cite. Once avoided emissions are deducted, a net increase of 0.47 to 1.8 billion tonnes per year remains, or 1.2 to 4.8% of global energy-related CO2 emissions in 2024.

The range is wide because it depends on assumptions about the size of the gains. The model describes an economic equilibrium, with no date, and is calibrated to the 2017 economy, with an electricity mix adjusted to that of 2024. In every case studied, as soon as AI brings any gain to the fossil sector, the emissions it enables exceed those it avoids in renewables. For a net decrease, renewables would need gains 4 to 5 times larger than those of the fossil sector. 2 of the authors declare a link with the Enabled Emissions Campaign, a non-profit research and advocacy project.

These emissions are poorly covered by emissions inventories. They appear in the oil producer's inventory, buried in the volumes it sells. They do not appear in the inventory of the company that sells it the tool: an AI provider counts its servers, their electricity and their manufacture, and no mandatory category of the GHG Protocol asks it to count what its customers produce with its models. The authors call on public policy to treat them as a separate category: enabled emissions.

In our reading, the case is similar to that of banks, which count the emissions of companies and of the projects they finance↗. The difference is that a financed emission is calculated pro rata to the amount financed, whereas an enabled emission is an additional amount estimated by a model. The question ties in with that of the place of AI in a Bilan Carbone®↗.

The experts surveyed also cite the self-driving car. They estimate that it would cut the cost of passenger transport by 27.1% in 2040 and raise demand by 21.8%, as median values. Tested with average AI electricity demand, this effect puts cumulative emissions between 2025 and 2040 more than 6% above the reference. It is the strongest of the indirect effects tested one by one.

Economists call this mechanism the rebound effect: what becomes cheaper is consumed more. The authors cite another example, the consumption of goods stimulated by advertising and online communication, and specify that they did not model it. The 11% figure therefore does not include it.

These effects are the least well understood. In the same overview, Valérie Masson-Delmotte writes, drawing on Arcep, the French electronic communications regulator, that the indirect effects of digital technology, including induction and rebound effects, are difficult to assess.

The 11% scenario contains an ingredient that is not a quantity of energy. In it, AI plays a large part in how information circulates; disinformation and polarisation weaken the consensus on climate targets, which governments set aside. The authors do not quantify this mechanism directly: it shapes the assumptions of the scenario, and they consider it hard to represent in a model.

They note a second obstacle, also political. Efficiency standards and rules on low-carbon electricity can be challenged on grounds of corporate and national competitiveness. The authors add that these rules also push towards more frugal systems.

Other work quantifies possible reductions. The Grantham Research Institute↗, with the economist Nicholas Stern, estimates that AI could reduce global emissions by 3.2 to 5.4 billion tonnes of CO2e per year by 2035, in only 3 sectors: electricity, transport and food. That would be more than the increase due to data centres and AI.

The 2 studies do not measure the same thing. The Grantham study adds up potentials, sector by sector. The February 2026 study runs a model in which efficiency gains trigger rebounds, and finds only a 1.4% reduction in its best scenario. They set the same condition: the Grantham study gives governments a decisive role and speaks of an active state; the February study, of strong governance and international cooperation.

The work cited here points to 4 decisions, each tied to one of the mechanisms.

The hub counts the electricity, water and CO2e of a question put to an AI. This work argues for counting direct and indirect effects together. In our reading, the footprint of AI has 3 tiers: what it consumes, what it makes possible and what it changes in public debate. The hub measures the first; according to these studies, the second could weigh more, and the third is not quantified.

AI can therefore amplify the climate crisis: in the least favourable of the scenarios studied, by 11% of cumulative CO2 emissions from 2025 to 2040. Part of this risk comes from what AI is used for, on top of its electricity consumption.