AI and Climate: The Efficiency Gains We Should Not Take for Granted
Artificial intelligence is often presented as a tool for accelerating the transition to a low-carbon economy. AI can optimize renewable energy production, improve electricity grids, reduce industrial waste, and make transport more efficient. At the same time, the rapidly growing energy consumption of AI itself has become a climate concern.
But there is another part of the story that is easy to miss.
A new study published in Nature / npj Climate Action suggests that the climate impact of AI cannot be understood simply by comparing the energy consumed by AI systems with the emissions they might help avoid. The researchers argue that AI is better understood as a bidirectional productivity amplifier: it can make both low-carbon and high-carbon energy systems more productive. And because the global economy is still deeply dependent on fossil fuels, those two effects are not symmetrical.
Uunder the study's modeled conditions, AI-driven productivity gains enable more CO₂ emissions than they avoid.
AI can make fossil fuels more competitive
Consider where AI can be applied in the energy sector. In renewable energy, AI can improve forecasting, optimize generation, support predictive maintenance, and help integrate renewable electricity into the grid. These applications can reduce emissions by extracting more useful energy from existing infrastructure.
But similar productivity improvements can be applied to fossil fuels. AI can improve exploration and extraction, reduce production costs, optimize refining, and make previously uneconomic resources commercially viable. In other words, AI does not inherently push the economy toward cleaner energy. It makes the underlying system more productive—whatever that system happens to be.
This distinction matters because fossil fuels still account for approximately 80% of global primary energy consumption. Renewable energy has been growing, but historically much of that growth has added to the total energy supply rather than directly displacing fossil fuels.
When AI improves productivity in both systems, the starting conditions therefore matter.
The surprising scale of the effect
The researchers modeled AI-driven productivity improvements across fossil fuel production, renewable energy generation, electricity infrastructure, and selected energy-intensive end uses. Rather than looking only at the direct emissions of computing infrastructure, they used a global computable general equilibrium model to capture economy-wide effects such as price changes, substitution between sectors, capital reallocation, and demand responses.
Under their parallel-adoption scenarios, AI-driven productivity gains increased global annual CO₂ emissions by 0.47–1.8 gigatonnes. That corresponds to approximately 1.2–4.8% of global energy-related CO₂ emissions in 2024.
The researchers also found that the emissions enabled by fossil-fuel productivity gains were substantially larger than the emissions avoided through renewable-energy optimization.
Most strikingly, they estimate that renewable productivity gains would have to be four to five times larger than fossil-fuel productivity gains to reach the modeled break-even point. Put differently, a 1% productivity improvement in fossil fuels would need to be accompanied by roughly 4–5% improvement in renewables to offset its modeled emissions impact.
The asymmetry was not limited to one particular set of assumptions. The researchers tested 64 combinations of fossil, renewable, and fuel-neutral productivity gains, as well as different model parameters and energy-mix assumptions. The direction of the result remained broadly consistent.
This is the rebound effect at a much larger scale
There is an important concept behind these results: rebound effects. An efficiency improvement does not necessarily translate into an equivalent reduction in resource consumption. If something becomes cheaper or more productive, we may simply consume more of it.
AI can take this mechanism further through what the researchers call induction effects. Productivity improvements can change the economics of production, affecting investment, resource allocation, prices, and the viability of previously uneconomic activities.
If AI makes oil or gas extraction cheaper, the result may not be less energy consumption. It can instead mean more economically viable supply, greater production and ultimately more fossil fuel consumption.
That is why measuring only the efficiency improvement of an individual AI application can give a misleading picture of its climate impact.
The datacenter footprint is only part of the story
Much of the discussion about sustainable AI focuses on the energy consumed by datacenters.
That is an important first-order impact. AI models require substantial computing infrastructure, electricity and cooling. But the study argues that this is only one layer of AI's climate impact. The researchers distinguish between:
-
First-order effects: the direct operational footprint of AI infrastructure, such as datacenter energy consumption.
-
Second-order effects: the consequences of applying AI to economic activities, including emissions enabled or avoided by those applications.
-
Higher-order effects: broader systemic dynamics such as rebound, induction, market responses and changes in economic activity.
In the modeled scenarios, enabled emissions were estimated at 0.6–2.4 gigatonnes of CO₂ annually. For comparison, the study notes that the IEA's estimate for datacenter emissions in 2025 is 0.18 gigatonnes. The authors emphasize that these figures come from different analytical frameworks and should not simply be added together, but the comparison illustrates the potential scale of indirect effects.
The implication is important: the climate impact of AI is not just about how efficiently AI runs. It is also about what AI makes possible.
"Green AI" is not automatically green
The findings challenge a common assumption in discussions about sustainable AI: that using AI to improve efficiency will naturally produce environmental benefits.
The study finds that fuel-neutral efficiency improvements—for example, improvements in electricity infrastructure and selected industrial and maritime applications—can reduce emissions. But in the modeled scenarios, they were not sufficient to reverse the effects of productivity gains in fossil fuel supply.
This does not mean that AI cannot contribute to decarbonization. It means that the application matters.
An AI system that improves renewable-energy forecasting is fundamentally different from one that makes fossil-fuel extraction more productive. Both may be highly efficient pieces of software. Both may reduce energy or operating costs within their immediate application. But their system-level climate consequences can be very different.
That distinction should become part of how organizations evaluate AI projects.
What should organizations do?
The study is primarily concerned with governance and economy-wide policy, but its findings also offer a useful principle for organizations adopting AI.
Instead of asking only how much energy this AI system consumes, we should also ask:
- What economic activity does this AI system enable?
- Does the productivity gain reduce resource consumption, or does it make additional consumption more economically attractive?
This leads to a broader assessment framework for AI sustainability.
Organizations should consider at least three levels:
1. Operational impact
How much energy and carbon does the AI system itself require?
2. Application impact
Does the application reduce emissions, enable emissions, or
potentially do both?
3. Systemic impact
Could the productivity improvement increase demand, expand production,
make previously uneconomic activity viable, or otherwise create
rebound or induction effects?
This is a much more demanding question than simply measuring the carbon footprint of an AI model—but it is also much closer to the actual environmental consequences of deploying AI at scale.
The policy challenge: don't optimize only the green side
The researchers propose five broad governance priorities. Among them are recognizing enabled emissions as a distinct category, assessing direct and indirect AI impacts together, combining renewable-energy optimization with constraints on AI-enabled fossil productivity, evaluating efficiency improvements according to their whole-system effects, and using carbon pricing as part of a broader policy portfolio.
Making green activities more efficient is not enough if we simultaneously make fossil activities more efficient.
The study argues that market forces alone, under the conditions modeled, reinforce rather than displace carbon-intensive systems. AI productivity gains sustain fossil-fuel supply economics, increase fossil consumption and amplify economic activity within fossil-dependent systems.
That does not make AI inherently bad for the climate. Rather, it makes AI a technology whose environmental impact depends heavily on where and how its productivity gains are applied.
A more mature conversation about sustainable AI
Technology does not operate in isolation from the economic system around it. A more efficient technology can reduce environmental impact—or it can accelerate the system that causes the impact in the first place.
AI makes this particularly visible because it is a general-purpose technology. The same capabilities—prediction, optimization, automation and pattern recognition—can be applied to renewable energy, electricity grids, logistics, manufacturing, oil exploration or almost any other economic activity.
The study's model is not a forecast of exactly how much CO₂ AI will produce. It is a comparative-static model designed to examine the direction and approximate scale of economic effects under specific assumptions. The authors explicitly note that it does not capture time-dependent dynamics and that real-world magnitudes may differ. It also models CO₂ from fossil-fuel combustion rather than the full range of greenhouse gases.
If AI makes everything more productive, what happens when "everything" still includes a fossil-fuel-dependent economy?
The answer cannot be found by looking at datacenter electricity consumption alone. Sustainable AI requires us to examine the full chain from computation to application to economic response.
AI may help accelerate the energy transition. But that outcome is not automatic. It depends on what we choose to optimize—and what we choose to constrain.
As the study concludes, AI's climate trajectory will depend on whether coordinated interventions can redirect its productivity gains toward displacement of carbon-intensive systems rather than reinforcement of them.
What does this mean for companies and their sustainability efforts?
The research shows that AI can simultaneously improve the efficiency of low-carbon activities and make fossil-fuel-based activities more productive. In the modeled scenarios, the latter effect dominates: AI-driven productivity gains increase net emissions unless gains in renewable energy substantially outpace gains in fossil-fuel productivity.
This has several practical implications:
-
Move beyond the carbon footprint of AI
Companies should look beyond the direct energy consumption and carbon footprint of AI systems to consider their wider economic and environmental effects. Assessments should consider how AI-driven productivity may affect production, consumption, resource use, and emissions, including potential rebound and induction effects. -
Ask what productivity is being increased
AI should not be considered environmentally beneficial simply because it makes a process more efficient. Companies should assess what activity AI makes more productive and whether it helps displace carbon-intensive activities or instead reinforces them. -
Include AI in existing sustainability governance
AI sustainability should be integrated into existing environmental and sustainability governance rather than treated solely as an IT responsibility. AI investments and use cases should be assessed for their direct and indirect environmental impacts alongside the company's broader sustainability objectives. -
Be more careful with "AI for sustainability" claims
Companies should avoid describing AI applications as sustainable simply because they improve efficiency or are intended to reduce emissions. Sustainability claims should distinguish between efficiency improvements, avoided emissions, and emissions enabled by increased productivity, and consider the net impact of the application.
About this research
This article is based on the research paper AI-driven productivity gains enable more CO₂ emissions than they avoid in a global energy–economy model, published in npj Climate Action on 4 August 2026. The study was authored by Will Alpine, Nathan Geldner, Holly Alpine and Maksym G. Chepeliev.