Does AI pollute more than aviation?

The comparison circulates in 2 forms: "digital technology pollutes more than aviation" and, since the arrival of ChatGPT, "AI pollutes more than aviation". The first can be verified for aviation's CO2 alone; with its contrails, the gap narrows. The second cannot be verified today, nor in 2030 in the GreenIT projection. At the scale of one person, a long-haul flight weighs thousands of times one year of everyday AI use.

Flights worldwide emitted 942 million tonnes of CO2 in 2024, according to ATAG, the organisation that brings together the aviation sector, compared with 882 million in 2023. This is about 2% of the CO2 emitted by human activities. For AI, the only worldwide life cycle estimate we have found is that of GreenIT: 99 million tonnes of CO2e in 2025, counting the manufacture and electricity of servers, storage, network and data centres. AI therefore emits about a tenth of the CO2 of aviation.

The International Energy Agency (IEA) counts differently: 180 million tonnes of CO2 in 2024 for the electricity of all the world's data centres, of which AI is only a part. This wider scope, which includes all the other digital uses hosted in these centres, represents about a fifth of the CO2 of aviation.

The 942 million tonnes count only the CO2 from engines. An aircraft at altitude also leaves contrails, which form clouds, and nitrogen oxides. The reference study on these effects, published by Lee and co-authors in 2021, estimates that they make up 2 thirds of the radiative forcing of aviation in 2018: the total radiative forcing of aviation was about 3 times that of its CO2 alone. This ratio is not a CO2e equivalence factor, which depends on the period over which one counts, but it shows that a comparison in CO2 alone underestimates the weight of aviation.

At the scale of a single flight, ADEME counts these effects in the Base Carbone: for a long-haul flight, its factor with contrails is 1.8 times the factor without contrails. The worldwide comparison on this page, which sets aviation's CO2 alone against the life cycle of the servers, storage, network and data centres of AI, is therefore favourable to aviation.

The sentence "digital technology pollutes more than aviation" comes from another scope: telephones, computers, televisions, networks and all data centres. GreenIT estimates it at 1,832 million tonnes of CO2e in 2023, 3.4% of global emissions, about 2 times the CO2 of aviation. Computer centres account for 22.8% of it. With its non-CO2 effects, about 3 times its CO2 alone according to Lee and co-authors, aviation would weigh on the climate as much as digital technology, or even more. Other estimates range from 1.4 to 3.9% of global emissions depending on the scope; the chart of digital technology in global emissions compares them.

GreenIT projects 636 million tonnes of CO2e for AI in 2030, 6 times more than in 2025. It assumes servers used 24 hours out of 24 and electricity that does not decarbonise, and exceeds the IEA projections for all data centres together. The IEA, for all data centres, forecasts a peak of about 320 million tonnes of CO2 around 2030 in its base case scenario, and a plateau of about 475 million at the start of the 2030s in its high-growth scenario. ATAG does not publish a projection of aviation emissions for 2030; they increased by 7% between 2023 and 2024.

Between 2025 and 2030, in this projection, the gap narrows: AI emissions grow there by about 45% a year; the analysis " How could AI amplify the climate crisis? " examines what would make it weigh more.

A return trip Paris - New York, about 11,700 km, emits 2.1 tonnes of CO2e per passenger with the ADEME Base Carbone factors, contrails included, and 1.1 tonne without them. This is about 25% of the annual carbon footprint of a French person, 8.3 tCO2e. A one-way flight Paris - Marseille emits about 150 kgCO2e with contrails.

On the other side, one year of everyday personal AI use weighs about 340 gCO2e for the electricity of the servers, according to the AI Impact Database. According to our calculation, it would take about 6,100 years of this use to equal a return trip to New York. The heaviest profile in the Database, a developer who hands tasks to code agents every day, reaches 180 kgCO2e a year: slightly more than a one-way flight Paris - Marseille, 11 to 12 times less than a return trip to New York.

For us at Projet Celsius, the comparison with aviation reads differently depending on the scale. For one person, a transatlantic return flight weighs more than a year of AI use, even for the most intensive profile in the Database, with only server electricity counted. Worldwide, AI emits about one tenth of the CO2 of aviation today, but its emissions are growing faster: in GreenIT's projection, about 45% a year between 2025 and 2030, against 7% for aviation between 2023 and 2024. We will update this page with each new estimate from GreenIT, the IEA and ATAG.