How Can AI Increase or Reduce Carbon Emissions?

Artificial intelligence does not have one fixed climate effect. Its footprint depends on the electricity used to build and run models, but also on what those models help people and companies do. An AI system can reduce wasted electricity in a grid, or help an oil producer find and extract additional fuel.

Where do AI emissions begin?

The most visible source is computing infrastructure. Training and operating models requires servers, cooling equipment, networking and data storage. The resulting emissions depend on total electricity consumption, the power grid’s fuel mix, hardware manufacturing and how intensively equipment is used.

A model running on a grid dominated by coal or gas has a different operational footprint from the same workload supplied by low-carbon electricity. Efficiency improvements can lower the energy needed for each task, although falling costs may encourage more use and offset part of the saving.

How can AI avoid emissions?

AI can forecast wind and sunlight, schedule batteries, identify faults before equipment fails and balance supply with demand. In buildings and factories, it can adjust heating, cooling and production around actual conditions instead of fixed schedules.

These benefits must be measured against a credible baseline. A forecast system avoids emissions only if it causes a real operational change, such as replacing gas-fired generation, reducing curtailment or preventing wasted energy. A laboratory accuracy score alone does not establish a climate saving.

How can the same tools enable more emissions?

Productivity gains can expand carbon-intensive activity. Oil and gas companies use machine learning to interpret seismic data, select well locations, predict equipment problems and improve recovery from reservoirs. If those improvements lower production costs or increase recoverable reserves, more fuel may reach the market.

This is an indirect effect. The computing system may consume relatively little electricity while enabling a much larger quantity of fuel to be extracted and burned. Similar effects can appear in aviation, shipping, mining and manufacturing when optimization lowers costs enough to stimulate additional demand.

Why is efficiency not automatically a climate benefit?

Efficiency measures output per unit of input, not total output. A process can use less energy for each product while producing so many additional units that overall consumption rises. Economists call this response a rebound effect.

The relevant question is therefore not only whether AI makes a task more efficient. Analysts must also ask whether prices fall, demand rises, production expands or investment shifts toward a higher-emitting activity.

How should a climate claim be tested?

• Define the activity changed by the AI system and the period being measured.
• Compare it with a realistic alternative, not with doing nothing unless that is genuinely the likely baseline.
• Include model operation, hardware and relevant changes in production or consumption.
• Separate measured results from modeled estimates and company projections.
• Check whether the benefit survives changes in electricity mix, adoption rate and market demand.

What information is still missing?

Technology companies disclose limited detail about energy use, while customers rarely report how AI changes output across their operations. Energy producers may announce cost savings without showing whether total extraction increased. These gaps make direct computing emissions easier to count than consequences elsewhere in the economy.

A complete assessment needs both ledgers: the resources consumed by computing and the emissions increased or avoided by its applications. Until deployment and production data improve, large claims in either direction should be read as scenarios with stated assumptions, not as universal properties of AI.