Is AI Bad for the Environment?

The short answer

AI does have a real environmental cost, and it is not trivial. Almost all of it is electricity: the power drawn by the data centres that train these models and then run them every time somebody asks for something, plus the water many of those buildings use for cooling and the emissions tied up in manufacturing the hardware. What is genuinely unclear is the size of that cost. The published estimates disagree with each other by a wide margin, for reasons worth understanding, and the decisions that actually move the needle sit with the companies building this and the electricity grids powering them rather than with you.

Where does the cost actually come from?

There are four places, and confusing the first two is the source of most of the muddle you will read online.

Training. Before a model can answer anything, it has to be built. That means running an enormous amount of computation across a very large number of specialised chips for a sustained period. It is expensive, it is energy hungry, and it happens once per model. When you see a dramatic headline figure attached to AI, it is often a training figure, which says nothing about what happens when you use the finished product.

Inference. This is the model actually doing something: answering your question, drafting your email, summarising your document. Each individual request costs a small fraction of what training costs. But inference happens constantly, across an enormous number of users, forever, for as long as the model is in service. Small multiplied by vast is not small. Over the lifetime of a popular model, the running cost can outweigh the cost of building it.

Cooling. Chips working hard produce heat, and heat has to go somewhere. Some facilities use water evaporation to shed it, which consumes fresh water, sometimes in places where water is already under strain. Others use air or closed loop systems, which use less water but more electricity. The trade off is a design choice that varies from building to building.

Hardware. The chips themselves have to be mined for, manufactured, shipped and eventually disposed of. This is a real cost that most estimates ignore entirely, partly because it is genuinely hard to measure.

Why do the numbers you see vary so wildly?

This is the most useful thing to understand, because it is what lets you judge any claim you meet in future.

Almost nobody publishes the raw data. The companies running these systems know their own consumption in detail and disclose very little of it. So most public figures are estimates built by outside researchers from whatever is visible: chip specifications, reported facility capacity, and reasonable guesses about efficiency. Change any one of those assumptions and the answer moves a lot.

A query is not a fixed unit of work. Asking for a one line answer and asking for a long piece of reasoned analysis are wildly different amounts of computation, even from the same model. Generating an image or a video is different again. A single number attached to the phrase "one AI query" is averaging across things that have almost nothing in common.

Models differ from each other enormously. A small model designed to run cheaply and a very large frontier model are not comparable. A figure carefully calculated for one specific system routinely gets quoted as though it applies to all AI, everywhere, forever. It does not.

The boundary keeps moving. Does the estimate include cooling? Manufacturing the chips? The share of the building's overhead not directly doing computation? Training, spread across every use? Reasonable people draw that boundary in different places and arrive at very different totals honestly.

Everything is changing fast. Chips get more efficient, models get optimised, and the total amount of usage keeps climbing. A figure that was accurate when it was calculated may be badly wrong within a year, in either direction.

Why does the electricity mix matter more than the query count?

Here is the part that gets lost. The same request, running on the same hardware, has a very different carbon footprint depending on which grid the building is plugged into and what that grid happens to be burning at that moment. A data centre on a grid running mostly on wind, solar, hydro or nuclear produces a small fraction of the emissions of an identical facility on a grid leaning on coal.

Which means the highest leverage question is not "how many times did people use this today". It is "how is the electricity being made, and is new demand being met by clean generation or by keeping older fossil plants running longer". That is a question about energy policy, siting decisions and grid investment. It is genuinely important, and it is not answered by anybody using a chatbot less.

Water works similarly. The same cooling load matters far more in a water stressed region than in a wet one. Where a facility is built can matter more than how hard it is worked.

Does your own usage matter?

Honestly? Not very much, in isolation. Your personal share of all this is small, in the same way your personal share of the internet's electricity is small. Nobody should feel guilty for using a tool to write a difficult email.

That said, "your individual share is small" is not the same as "nothing matters". Two things are true at once: individual restraint will not solve this, and being thoughtful about what you use and why is still a perfectly reasonable way to behave. The useful version of thoughtfulness is not agonising over each request. It is noticing when you are using something powerful for something trivial, or generating the same thing again and again out of habit rather than need.

If you want a clearer picture of what these systems are actually doing when you prompt them, our free Introduction to Generative AI course covers the mechanics in plain English, and Understanding AI Loops explains why a system that keeps calling itself does far more work than a single question does.

What actually reduces the impact?

In general terms, and without pretending any of it is dramatic:

Match the tool to the job. Reaching for the largest, most capable model to rephrase a sentence is like taking a lorry to fetch a pint of milk. Most tools now offer a lighter option, and for small jobs it is usually both faster and cheaper.

Stop regenerating out of habit. Hitting regenerate repeatedly and hoping for something better costs a full run every time. A better prompt on the first attempt costs one. This is the single biggest difference most people can make to their own usage, and it happens to save you time as well.

Batch your work. Ten related questions in one focused session, with the context supplied once, is less wasteful than ten scattered conversations that each start from nothing.

Be sceptical of automation that runs unattended. A scheduled job that quietly calls a model over and over, whether or not there is anything to do, will out consume your entire manual usage without you noticing.

How should you read the next headline you see about this?

When a striking figure crosses your feed, ask a short list of questions before repeating it.

WHO measured it?     The operator, or an outside estimate?
WHAT was measured?   Training, or running? One model, or all AI?
One request, or a whole facility?
WHICH boundary?      Does it include cooling, manufacturing,
idle overhead, the training cost spread
across uses?
WHERE?               Which grid, and how is it powered? A figure
with no location attached is missing the
most important variable.
WHEN?                Efficiency and usage both move fast. Old
figures go stale in both directions.
COMPARED to what?    Comparisons to cars, kettles or flights are
chosen for effect. Ask who picked the
comparison and why.

A claim that survives those questions is worth taking seriously. Most of what circulates does not, in either direction: the reassuring numbers are as unreliable as the alarming ones.

The honest summary is that this is a real cost, poorly measured in public, growing quickly, and shaped far more by how the industry is powered than by how often any one of us types a question. Holding that view is less satisfying than outrage or dismissal, but it is the one the evidence currently supports.

If you want to think more carefully about where AI helps and where it should be treated with caution, AI Safety and Verification is a good next step, or you can browse all our free courses and start wherever you like.