Is AI bad for the environment? The honest answer

Yes, AI can be bad for the environment. Training and running models use electricity; data centres may consume water for cooling; and making chips and buildings requires materials and creates emissions. At large scale, those costs matter. But “AI” covers everything from a tiny classifier to a frontier model generating video, so there is no single footprint.

The defensible conclusion is conditional: an AI system is environmentally harmful when its full costs are high relative to the value it creates and when cleaner, smaller or non-AI alternatives could do the job. AI can also help operate electricity grids, find methane leaks and reduce waste. Those benefits must be demonstrated, not assumed. This is why both “AI is destroying the planet” and “AI will solve climate change” are too simple.

Electricity is the clearest pressure

AI runs in data centres, where accelerators perform the calculations behind training and answers. The International Energy Agency estimated that all data centres—not AI alone—used about 415 terawatt-hours of electricity in 2024, around 1.5% of the global total. Its base case reaches roughly 945 TWh in 2030, with AI the most important driver of growth alongside other digital services.

Those are projections, not a meter reading from the future. Uptake, hardware efficiency, model design and limits on grid connections could move the total substantially. The IEA therefore publishes several scenarios.

A single ordinary text exchange is much smaller. Recent public estimates put a typical prompt at a fraction of a watt-hour, while long context, reasoning, search and generated images can require much more. Small per-use figures and fast-growing system demand can both be true: efficiency per answer can improve while total use rises faster.

Carbon depends on where and when power is made

Electricity use is not the same as carbon emissions. A kilowatt-hour supplied by a coal-heavy grid has a different operational footprint from one supplied by wind, solar, hydro or nuclear power. Location and time therefore matter, as do the rules used to account for renewable-energy contracts.

The IEA’s physical-grid analysis says fossil fuels still supply a substantial share of data-centre power. It projects electricity-related data-centre emissions peaking at about 320 million tonnes of CO2 around 2030 in its base case. That remains less than 1% of global CO2 emissions, but data centres are also one of the sectors in which emissions rise to 2030 in that analysis.

Company claims need careful reading. “Matched with renewable energy” may mean a company bought enough clean electricity over a year, not that a particular server used carbon-free power in the hour it answered. Procurement can support new generation, but it does not make geography and timing disappear.

Water is local, variable and poorly disclosed

Some data centres evaporate water to carry heat away. Electricity generation can consume water too, creating an indirect footprint even at a site that uses little water for cooling. The result changes with weather, cooling design and the local power mix.

Berkeley Lab modelled average on-site water consumption across US data centres at just over 0.36 litres per kilowatt-hour in 2023. That is a fleet average, not a universal rate. A cool-climate facility using outside air can be very different from an evaporatively cooled facility in a hot region.

Per-message claims vary even more. A widely shared academic estimate linked a 500 ml bottle to roughly 20–50 language-model responses; later operator figures for median text prompts were much lower. Their models, locations and accounting boundaries differ. Our guide to how AI uses water explains why no bottle comparison applies to every message.

Hardware has a footprint before AI is switched on

Operational energy and water are only part of the life cycle. Accelerators require mined and processed materials, semiconductor fabrication, transport and eventually disposal. Data-centre construction uses steel, concrete, cables and backup equipment. Manufacturing can also require ultrapure water and produce greenhouse gases.

Public per-query estimates seldom allocate all of those impacts. There is no agreed, audited method for dividing a chip’s manufacturing footprint among training, inference and every other job performed during its useful life. Leaving hardware out makes a number incomplete; allocating it with speculative utilisation assumptions can create false precision.

The sensible response is to state the boundary. An inference-energy estimate can answer “roughly how much operating electricity did this answer require?” It cannot, by itself, answer “what was the complete environmental impact of this AI service?” Our methodology follows that rule: Treechat’s receipts are estimates based on model and token use, data-centre overhead and an average grid factor, not full life-cycle measurements.

AI can help, but avoided impact must be proved

There are credible beneficial uses. The IEA describes applications in power forecasting, grid operation, industrial optimisation and detecting methane leaks. Better forecasts can help integrate variable renewable generation; faster leak detection can prevent emissions; and building controls can reduce wasted energy.

Yet a possible application is not an achieved saving. The relevant comparison includes the AI system’s own footprint, deployment at scale, human behaviour and rebound effects. A route optimiser that saves fuel may be beneficial. If lower costs lead to much more driving, the net result can shrink or reverse. A model that discovers a lower-carbon material still needs laboratory validation and adoption.

This is the key distinction in asking whether AI is good for the environment: capability is not outcome. Look for a measured baseline, a credible counterfactual, a stated system boundary and results over time. Be wary when a provider counts every possible saving but omits the computing required to produce it.

How bad is a particular use?

Start with the task. A specialist model doing one classification can be far lighter than a general-purpose generator. In experiments across common machine-learning tasks, Luccioni, Jernite and Strubell found generative systems were substantially more energy-intensive than task-specific models performing comparable functions. Their benchmark results are not universal product measurements, but they show why model and task choice matter.

Next ask about volume and mode. A few short text answers are not equivalent to millions of long outputs or repeated image and video generation. Then ask where the service runs, how efficiently the facility uses power and water, and what the provider actually discloses.

Finally, consider alternatives. A normal search, calculator, local script or existing document may solve the problem with less computation. For work that genuinely benefits from AI, use the smallest capable model, keep context focused and avoid generating disposable variations. See our fuller myth audit of how bad AI really is.

What a responsible reader can do

You do not need a perfect personal carbon calculator to make better choices. Reserve heavy reasoning, agents, images and video for work that needs them. Use short, clear prompts and stop when the answer is good enough. Choose providers that disclose methods rather than presenting precise numbers without boundaries.

Individual restraint is not a substitute for infrastructure policy or corporate reporting. Operators control facility location, cooling, hardware utilisation and energy procurement. Governments and regulators shape grid planning, water permits and disclosure. Large customers can ask suppliers for workload-specific energy, water and emissions data.

Treechat’s approach is deliberately narrower: route everyday questions to smaller models, show an estimated receipt and fund one newly planted tree each month for every paid member. That planting does not erase the computation. It is a separate contribution, while model choice tackles the operational estimate. The broader explanation of how AI affects the environment is still the better frame than any green or guilty label.

Frequently asked questions

Is one AI question bad for the environment?

One short text question has a small operational footprint, but it is not zero. The model, answer length, data-centre efficiency and electricity mix determine how small. The larger concern is repeated use across millions of people and more intensive modes.

Is AI worse than Google Search?

There is no timeless ratio. Search and AI products change, some searches now include generated answers, and published estimates use different years and boundaries. Compare a defined task using current, like-for-like data.

Can renewable energy make AI green?

Cleaner electricity can greatly reduce operational emissions. It does not remove water use, hardware manufacturing, construction or every local grid effect, and annual renewable matching is not the same as hourly carbon-free operation.

Should I stop using AI?

Not necessarily. Use it where it produces enough value to justify the resources, and prefer a smaller capable model. The better question is not whether all AI is good or bad, but whether this system is a proportionate way to do this task.