Artificial intelligence promises more efficient electricity grids, more renewable energy generation and lower energy consumption. But the same technology is also making oil, gas and coal extraction more efficient. New research in npj Climate Action shows that the latter effect could be significantly stronger than the former.
When discussing the climate footprint of artificial intelligence, the conversation usually starts with data centers. Training and running AI models requires huge amounts of electricity, but this is offset by the potential benefits: AI could manage electricity grids more efficiently, forecast the output of solar and wind farms and reduce consumption.
Researchers in npj Climate Action offer a much broader view. In the paper “AI-driven productivity gains enable more CO₂ emissions than they avoid in a global energy-economy model,” published on August 4, 2026, they ask a different question:
What if AI makes not only green energy but also the fossil fuel industry more productive?
The authors see artificial intelligence as a “two-way productivity amplifier.” The technology has no inherent bias toward a low-carbon economy.
In renewables, AI can improve production forecasts, preventive maintenance, generation efficiency, and grid integration. But its applications now extend across the fossil fuel chain. Algorithms can increase extraction efficiency and reduce marginal production costs and operational risk. The result is that resources that would otherwise be uneconomical can become profitable to exploit.
AI can, for example, improve upstream oil production decisions so that more output is extracted from the same amount of capital, labor, and energy. The authors also give a counterexample: algorithmic detection of methane leaks can reduce emissions from oil and gas operations. That is, even within the fossil fuel industry, AI has two opposing effects - it can make a unit of production cleaner, but at the same time make production cheaper overall and increase the amount of fuel extracted.
Goldman Sachs, for example, estimates that improvements thanks to AI could increase US shale oil reserves by 8-20%, reduce drilling costs by about 30%, and lower marginal production costs by approximately $5 per barrel. This could potentially make projects that were previously on the verge of profitability economically viable.
Data from operators themselves confirm this effect in two ways. On the one hand, AI allows resources that were previously uneconomical to be brought into production – the so-called extensive margin. On the other hand, the technology increases production and reduces costs in fields that are already being exploited – the so-called intensive margin.
Ultimately, it turns out that more efficient extraction could mean more, not less, emissions – between 470 million and 1.8 billion tons more CO₂ per year.
To measure this effect, scientists use a global general economic equilibrium model that takes into account changes in production, prices, consumption and substitution between sectors. The results are scenarios for the direction and approximate magnitude of the effect, not a forecast for a specific future year.
And the result is significant. If AI enters renewable and fossil energy in parallel, the model shows an increase in global annual emissions of 0.47–1.8 gigatonnes of CO₂, or 1.2–4.8% of global energy-related emissions in 2024.
Even more interesting is the comparison with data centers. The modeled additional emissions due to higher productivity in fossil fuels are 0.6–2.4 gigatonnes per year – 3.3 to 13.3 times the International Energy Agency’s estimate of emissions from data centers in 2025. The two numbers can’t simply be added together, the authors warn, but the comparison shows the scale of the hitherto overlooked effect.
So AI’s biggest climate problem may not be the electricity that AI itself consumes, but the economic activity it makes cheaper and more productive.
The study finds a strong asymmetry between the two energy worlds. For every 1% increase in fossil fuel efficiency, it takes about 4-5% in renewables to reach net zero emissions. And the reason is important: the main problem is not that AI makes a gas or coal power plant more efficient. The model shows that the decisive factor is the extraction of oil, gas and coal. When AI lowers costs and increases efficiency there, the economically affordable supply of fossil fuels expands.
The authors test the relationship in 64 scenarios. When AI is deployed in parallel, emissions increase in all of them. Even a 15-20% improvement in efficiency in key renewable technologies, combined with just a 2-5% improvement in fossil fuel extraction, still leads to an increase in emissions. This is a classic economic rebound effect on a massive scale: technology allows something to be produced more efficiently, it becomes more competitive, and ultimately the economy uses more of it.
Scientists are also testing scenarios with a carbon price.
At $80 per ton of CO₂, the additional emissions are still about three times the avoided emissions, and the net result remains plus 0.7 gigatonnes. Even at $308 per ton, the model still shows about 0.1 gigatonne of net increase.
Improvements in electricity grids and efficiency in industry and transport mitigate the effect, but they also do not reverse the trend. With the parallel implementation of AI, emissions in the model even grow faster than GDP – Instead of technological progress decoupling economic growth from carbon emissions, it is increasing carbon intensity.
Artificial intelligence is neither “green” nor “dirty” in itself. It is a productivity-enhancing technology. And when implemented in a global economy where over 80% of primary energy still comes from fossil fuels, it has a strong chance of making that very system more efficient.
The study has important limitations. The model is static, and the results are not a deterministic forecast. The real economy will change along with technology, policies, and the energy mix.
But the main message remains clear: Technological progress does not guarantee decarbonization. AI can simultaneously detect a methane leak and make the next barrel of oil cheaper to extract. And if the second effect prevails, the climate balance becomes negative.
This shifts the debate to a much more difficult question. Not just how much electricity will AI consume, but what kind of economy will it make more productive.