Is AI good for the environment?

AI can be good for the environment when it produces verified reductions in energy, emissions, water use or waste that exceed the footprint of building and running it. It can improve weather and power forecasts, detect methane leaks, optimise equipment and help researchers analyse complex data. None of those benefits makes AI environmentally good by default.

The correct test is a comparison: what happened with AI versus what would probably have happened without it? That comparison must include the AI system’s electricity, water and hardware impacts, along with rebound effects. A promising demonstration is not the same as a durable saving at scale.

It should also ask who benefits and where costs land. A global saving can coexist with greater pressure on the grid or water supply of the community hosting the computing.

Where AI can plausibly help

Electricity systems constantly balance changing supply and demand. Machine-learning forecasts can help operators anticipate wind, solar and load, while optimisation can support maintenance and network operation. Industry can use anomaly detection to identify failing equipment or wasted energy.

The IEA’s Energy and AI report describes applications including faster methane-leak detection, power-plant optimisation and transport efficiency. These are credible mechanisms: finding a leak sooner can prevent emissions, and better controls can reduce fuel.

AI can also help process satellite imagery, map habitats and explore candidate materials. But “AI could” signals potential, not measured net benefit. Performance in a benchmark may not survive integration costs, poor data or human decisions in the real world.

The counterfactual is the heart of the claim

Suppose an AI building-control system reports a 10% energy saving. The meaningful baseline is not an uncontrolled building if competent conventional controls would already save most of that energy. The comparison should use the best realistic non-AI alternative.

Next include the whole intervention: sensors, networking, model training, inference and staff. A small operational footprint may be easily outweighed by avoided energy. A computationally heavy system delivering a marginal improvement may not be.

Finally, specify time and scope. A pilot over mild weather does not establish annual results. A local reduction can shift costs elsewhere in a supply chain. Strong studies predefine metrics, report uncertainty and monitor whether savings persist.

Rebound can shrink an efficiency gain

When optimisation makes a service cheaper or faster, people may use more of it. More efficient routing could encourage additional deliveries; cheaper generation can produce a flood of disposable media. This rebound effect does not mean efficiency has no value. It means total resource use must be tracked alongside use per task.

The same applies to computing itself. Hardware and software continue to improve, yet the IEA projects all data-centre electricity rising from about 415 TWh in 2024 to roughly 945 TWh in 2030 in its base case. AI is a major driver, although data centres run many non-AI services.

A provider celebrating lower energy per response should also disclose absolute energy and workload growth. Without both, readers cannot tell whether efficiency is reducing total impact or merely enabling expansion.

AI has costs even in a beneficial project

Training and serving models uses electricity. Cooling may consume water directly, while power generation adds an indirect water footprint. Chips and data-centre construction have embodied impacts. Our guide to how AI affects the environment traces those stages.

Cleaner electricity can lower operational carbon, but renewable matching does not remove materials or all local grid effects. A water-saving application may still run in a water-stressed data centre. A project can be net beneficial while retaining these costs; the claim simply needs to count them.

This is also why tree planting should not be called an eraser. Every paid Treechat member funds one newly planted tree each month as a separate contribution, while smaller models reduce estimated operational energy. Its methodology describes the estimate and its boundaries rather than declaring the chat impact-free.

How to recognise a credible green application

Look for a named environmental problem and a measurable outcome, not a broad promise to “unlock sustainability”. Ask whether the system is already deployed, who measured the result and over what period. Check the comparison and whether uncertainty is reported.

The claimed benefit should use the same boundary discipline as the cost. Avoided emissions based on a physical reduction differ from hypothetical savings if every customer adopted a feature. Company-wide enabled savings should not be confused with reductions in the company’s own footprint.

Independent replication is valuable, especially when the provider selling the AI also calculates the benefit. So is openness about failure. Some applications will not outperform simple rules, better insulation, ordinary software or direct policy.

Reporting failed pilots helps other organisations avoid repeating their computing and integration costs, and makes the successful cases more credible.

Use AI where it has leverage

For an individual, high-leverage uses include avoiding a wasted journey, repairing rather than replacing an item, understanding an energy bill or finding a relevant public service. Whether AI is necessary depends on the task; a direct source or calculator may be better.

Organisations should begin with the environmental decision, establish the baseline and choose the simplest capable tool. Pilot with clear stop conditions. Measure the AI workload and outcome, then scale only if the net result is positive.

The answer to whether AI is bad for the environment and the answer here are compatible. AI is physical infrastructure with genuine impacts, and it is also a tool that can reduce larger impacts. Neither side cancels the other. What matters is demonstrated net effect, scale and alternatives—not a green label attached to the technology.