When climate negotiators gather in Antalya, Türkiye, from 9 November for COP31, they will arrive with a new and awkward item on the agenda: artificial intelligence. The Turkish presidency has already launched an "Antalya AI Pledge" calling for AI that is consistent with climate goals, and the reason the subject has forced its way in is that the public mood has curdled. A CNN poll in September found nearly 90 percent of Americans are more worried than excited about the data centres powering the AI boom. The trouble for the summit is that the evidence does not point cleanly in either direction.
Start with the case for worry, because the numbers are large. The International Energy Agency says electricity use by AI data centres jumped 50 percent last year and could climb from about 415 terawatt-hours in 2024 to as much as 1,700 by 2035. Because data centres cluster geographically, the strain is local and sharp: in the United States, which hosts nearly half of global consumption, wholesale capacity prices in the PJM grid region leapt from under 29 dollars per megawatt-day in 2024 to over 329 this year, and industrial power rates have risen 26 percent in Ohio and 31 percent in Pennsylvania against 7 percent nationally. Water is the quieter cost. Meta alone used roughly 21 billion gallons in 2024, up almost 120 percent in three years, most of it consumed indirectly at the power plants feeding its servers.
Then there is a finding that cuts deeper than any single data centre. A study published in a Nature Portfolio journal, highlighted this week by the watchdog group AlgorithmWatch, estimates that AI's use in the energy sector could add up to 1.8 gigatonnes of net carbon dioxide a year, close to 3 percent of global emissions. The reason is uncomfortable: the oil and gas industry was an early and enthusiastic adopter. Equinor has credited AI with helping discover 27 new deposits in a single year; Saudi Aramco says the technology helps keep its extraction costs near three dollars a barrel. By this accounting, AI is currently doing more to find fossil fuels than to replace them, an effect that today's climate and ESG rules barely capture.
And yet the opposite case is also real. The same IEA reckons that if AI were deployed across transport, industry and buildings, it could save 13.5 exajoules of energy by 2035, roughly Indonesia's entire annual demand. AI weather models at the European Centre for Medium-Range Weather Forecasts now produce forecasts using about a thousandth of the energy of traditional methods, and Google says its flood-warning system can see seven days ahead in more than 150 countries. Climate TRACE, a nonprofit, uses AI and satellites to track emissions from some 745 million individual sources, giving cities an inventory they could never have built by hand. These climate-positive uses tend to run on small, specialised models, not the general-purpose giants behind the headlines.
So which AI shows up at COP31? That is really the question, and it has no single answer because AI is not one thing. The honest framing is that its climate effect is a choice, not a property: it depends on where the data centres are built, what powers them, how transparent the companies are forced to be, and, most pointedly, whether a model is tuned to optimise a wind farm or a drilling program. Negotiators cannot ban a general-purpose tool, but they can decide what to measure and what to reward. The uncomfortable lesson from this week's research is that, left to the market, AI has been drifting toward oil rather than away from it.