How Can AI Increase or Reduce Carbon Emissions?

Short answer: AI changes emissions through two different channels. Its computers, data centres, chips, and cooling create a direct footprint. Its applications also change the cost and productivity of activities across the economy. AI can avoid emissions by improving grids, transport, buildings, and materials, or enable more emissions by making fossil-fuel extraction, advertising, production, and consumption cheaper and faster.

Why the question is larger than data-centre electricity

Most discussion begins with the energy used to train and run AI models. That matters, but it is only the first layer. A routing model may consume electricity while reducing delivery kilometres. A drilling model may use little computing energy while helping extract far more oil. The indirect effect can exceed the software's own footprint in either direction.

A credible assessment therefore defines the system boundary: is it measuring one query, one model, a company, a sector, or the economy-wide response to higher productivity?

Four kinds of climate effect

Adding direct and embodied emissions gives the footprint of the AI system. Estimating avoided and enabled emissions asks what the system changes in the wider world. These numbers should not be mixed without explaining boundaries, timing, and the counterfactual.

Where AI can reduce emissions

• Electricity systems: forecasting demand and wind or solar output, detecting faults, scheduling storage, and optimizing power flows.
• Buildings: coordinating heating, cooling, ventilation, occupancy, and maintenance while preserving comfort.
• Transport and logistics: reducing empty journeys, congestion, fuel use, and unnecessary inventory.
• Industry: tuning processes, detecting defects, predicting maintenance, and reducing heat and material waste.
• Science and design: searching for catalysts, batteries, low-carbon materials, and more efficient components.
• Monitoring: identifying methane leaks, deforestation, equipment failure, and changes in land use.

Potential is not the same as realized reduction. A forecast only cuts emissions if operators can act on it, the displaced activity was carbon-intensive, and savings are measured against a credible baseline.

Where AI can increase emissions

Beyond its own electricity demand, AI can expand production. It can improve seismic interpretation, drilling placement, reservoir management, refinery operations, trading, and logistics in oil and gas. It can also make advertising more effective, accelerate product turnover, increase computing demand, and lower the cost of energy-intensive services.

The climate outcome depends on what productivity improves. A percentage gain in a very large high-carbon sector can outweigh a bigger percentage gain in a smaller clean sector.

The rebound effect

Efficiency lowers the resources required per unit of service, which often lowers cost. Demand may then rise and offset part of the expected saving. This is the rebound effect.

For example, AI may reduce fuel used per delivery, but cheaper delivery may encourage more shipments. It may make data-centre cooling more efficient while rapid growth raises total electricity consumption. Rebound can be partial, complete, or in rare cases greater than the initial saving. Its size must be measured rather than assumed.

Carbon accounting needs a counterfactual

“AI avoided one million tonnes of CO₂” is not directly observed in the way electricity consumption is. It compares the observed or modeled outcome with an alternative world in which the AI was not used. The result depends on what that alternative assumes.

1. Define the service: specify the outcome, such as tonnes of steel, passenger-kilometres, or heating comfort.
2. Build a baseline: describe what technology and behavior would reasonably occur without AI.
3. Measure all relevant changes: include computing, equipment, operational savings, additional demand, and displaced activity.
4. Use marginal emissions where appropriate: the generator responding to added electricity demand can matter more than the annual grid average.
5. State uncertainty: show ranges for adoption, rebound, grid mix, model lifetime, and future policy.

Why location and timing matter

The same computation can have different emissions depending on where and when it runs. A grid dominated by low-carbon power differs from one using coal or gas. Hourly marginal emissions can change even when the annual average stays fixed.

Workloads may be shifted to cleaner hours or regions when latency and data rules allow. But renewable-energy certificates do not automatically prove that a facility caused new clean generation or eliminated emissions at every hour.

Training versus inference

Training a frontier model is energy-intensive but occasional. Inference happens every time users or automated systems call the model and can dominate lifetime energy at large scale. Smaller specialized models, efficient hardware, batching, caching, and avoiding unnecessary generation can reduce energy per useful task.

Efficiency per query is still not enough. Total impact equals energy per task multiplied by the number and size of tasks, plus the broader changes those tasks cause.

A modeled global example

A 2026 global energy-economy study explored 64 scenarios for AI-related productivity. It estimated that gains in fossil-fuel production could outweigh modeled clean-energy benefits, adding about 0.47 to 1.8 gigatonnes of annual CO₂ under its assumptions. Read AI May Help Clean Energy, but Its Oil Gains Could Add 1.8 Billion Tonnes of CO₂.

Those figures are scenario outputs, not measured future emissions. The study depended on assumptions about sector productivity, adoption, energy demand, prices, and technology, and excluded some possible clean-technology breakthroughs. Its main lesson is structural: productivity affects both clean and fossil sectors, so benefits cannot be counted in isolation.

How to assess an AI climate claim

• What exact system, geography, period, and activity are included?
• Are operational and embodied computing emissions counted?
• What non-AI baseline is used?
• Are reductions measured, modeled, or only technically possible?
• Does the estimate include higher demand and rebound?
• Is the claimed saving gross or net of the AI footprint?
• Could the application lock in fossil infrastructure or accelerate clean substitution?
• Are uncertainty ranges and assumptions published?

What responsible deployment looks like

Organizations can measure energy and carbon by workload, prefer smaller adequate models, improve utilization, schedule flexible computing, procure credible low-carbon power, extend hardware life, and report embodied impacts. More importantly, they can prioritize applications with verified high climate value and restrict uses that expand high-carbon activity.

Policy also shapes the result. Carbon prices, clean-energy standards, methane rules, efficiency codes, and limits on fossil extraction determine whether productivity becomes lower emissions or simply more output.

The useful equation

A practical mental model is: net climate impact = direct and embodied AI emissions + enabled emissions − genuinely avoided emissions. Every term needs a stated boundary, counterfactual, time horizon, and uncertainty range.