Questions,
answered.

What a request costs, how we know, and what to do about it.

The basics

Are you against AI?
No, we use it too. A language model is a useful tool. The campaign is about one habit: reaching for it on reflex, for things you could do yourself, without noticing what it costs. Use it where it earns its keep, and skip it where it does not.
What is this campaign?
A quiet reminder, in the spirit of the old "please consider the environment before printing" line at the bottom of emails. We want people to pause for a second before asking a machine to summarise or generate a response, and to know what a request actually uses.
Is this site AI generated? How much did it cost, then?
Partly, and we would rather be honest about it than pretend it was free. The build was AI-assisted — a coding model helped wire up the templates and layout — but the site is mostly cached, templatised HTML, and almost all of the words you are reading were written and edited by a person rather than generated. The footprint is dominated by a modest run of reasoning-heavy coding requests. The rough working: a heavy request is around 30 watt-hours, so a few dozen of them come to something like one to two kilowatt-hours — under a kilogram of CO2 and a few litres of water, or less than a single hot shower. Hosting is a static site, so each visit after that is negligible. To offset it you would need people to skip a few dozen heavy requests, or a few thousand short ones — put another way, if this page talks a handful of readers into passing on one needless prompt a day, it covers its own cost within a couple of weeks.

The numbers

How much energy does one AI request use?
A typical short text answer is about 0.3 watt-hours on an efficient model, roughly what a desk lamp draws in a minute. A long answer from a large reasoning model can be thirty watt-hours or more, closer to a minute of an electric oven. The figure depends heavily on the model and the length of the reply.
What about water?
Data centres use water to cool their servers, and power stations use water to make the electricity that runs them. Counting both, an older model like GPT-3 worked out at roughly a 500 ml bottle for every ten to fifty short replies — tens of millilitres each, and it depends where the data centre sits. Efficient newer models use less.
Where do your figures come from?
Every number on the site is cited and linked, from Epoch AI, the International Energy Agency, and university research groups. They are estimates and vary by method. We try to state what each one includes — some count off-site electricity and the water used for cooling, while most leave out making the hardware and training the model — and we let you check the sources yourself.

Comparisons

Is printing a page worse than summarising it?
Usually yes. A sheet of A4 costs around five grams of CO2 in the making and several litres of water, nearly all of it in the paper itself. A short summary on an efficient model is a fraction of that. The campaign is not built on a single query beating a single sheet. It is built on scale, because we make billions of requests a day and many never needed making.

What to do

So should I never use AI?
No. Use it for the things it is genuinely good at. The five questions on the main page are the whole method: if you could read it, write it, or work it out yourself in about the same time, that is usually the better choice.
Does using it less actually matter at my scale?
One request is tiny. It is worth a thought because the habit is shared by hundreds of millions of people, and the total is now measured in hundreds of terawatt-hours. Small, repeated, reflexive choices are exactly the kind that add up.
What is the lowest-impact option?
Often it is neither printing nor prompting. It is reading the thing yourself, on the screen already in front of you.