# Green AI: a plain-English guide for everyone

> What using AI really costs in energy and water, which everyday habits matter and which don't, how to judge claims, and how to have your say.

Source: https://greenai.guide/guides/public/  
Updated: 2026-11-19

## In short

A single text question to an AI chatbot uses a small amount of energy, roughly a third of a watt-hour on the companies' own figures, and a fraction of a millilitre of water. Images and video use much more, and the sheer number of requests adds up. The habits that matter most are choosing the right tool, avoiding endless regeneration, and asking for honest information.

## Key points

- Google reports 0.24 Wh and 0.26 mL of water for a median Gemini text prompt; OpenAI says about 0.34 Wh for an average ChatGPT query. Both are self-reported.
- Generating images uses far more energy than handling text: in one study, over a thousand times more per request than classifying text.
- The International Energy Agency projects data-centre electricity use to more than double between 2024 and 2030, with AI the main driver.
- Per-prompt figures vary because companies measure different things; always ask what is included.
- You can have a say through your choices, your workplace or school, local planning decisions and public consultations.

## What is green AI, in plain English?

Green AI means getting the benefits of artificial intelligence while using less energy, water and other resources.
It covers two things: making AI itself more efficient, and using it sensibly, only where it helps.

It is not about telling people never to use AI. A 2026 UK green AI sandpit of mathematicians, computer scientists
and industry discussed this directly and concluded that doing AI responsibly is a more useful goal than refusing it.
There is a longer explanation in [what is green AI?](/articles/what-is-green-ai/)

## How much energy does using AI take?

For one ordinary text question, a small amount: around a third of a watt-hour on the companies' own figures. But
images and video take far more, and the number of requests worldwide is enormous.

Here is what has been published:

| Source | What it measured | Energy |
|---|---|---|
| Google (2025) | Median text prompt to Gemini apps | 0.24 Wh |
| OpenAI (2025) | Average ChatGPT query | about 0.34 Wh |
| Epoch AI (2025), independent estimate | Typical ChatGPT text query | about 0.3 Wh |

OpenAI compares its figure to what an oven uses in a little over one second, or a high-efficiency light bulb in a
couple of minutes. Epoch AI's estimate is about ten times lower than an older, widely quoted figure of 3 Wh, which
shows how quickly estimates change as hardware and software improve. Google reports that the energy of its median
prompt fell 33-fold over one year.

**Images and video cost much more.** In tests by researchers at Hugging Face and Carnegie Mellon University,
generating images used on average about 2.9 kWh per 1,000 images, compared with 0.002 kWh for sorting 1,000 pieces of
text into categories. That is more than a thousand times as much per request. Video generation is heavier still.

**The total is what worries energy planners.** The International Energy Agency (IEA) estimates that data centres used
about 415 TWh of electricity in 2024, around 1.5% of the world's electricity, and projects about 945 TWh by 2030.
It names AI as the most important driver of that growth.

## How much water does AI use?

Per text prompt, very little; in total, and in some places, a lot. Data centres use water mainly to keep computers
cool, and power stations use water to generate the electricity.

- Google reports 0.26 mL of water for a median Gemini text prompt.
- OpenAI says an average ChatGPT query uses about 0.000085 gallons, roughly one fifteenth of a teaspoon.
- Mistral AI, a French company, published a life-cycle assessment that counts more stages than either of those
  figures. It put a 400-token response (a few paragraphs) at 1.14 g of CO₂e (carbon dioxide equivalent) and 45 mL of
  water.
- Researchers Li, Yang, Islam and Ren estimated that training the GPT-3 model in Microsoft's US data centres could
  have directly evaporated about 700,000 litres of clean fresh water.

The big differences between these numbers come mostly from **what is counted**, not from one company being far
better than another. Where a data centre is matters too: water drawn in a dry region during a drought has a bigger
effect than the same amount where water is plentiful.

## Why are the numbers so uncertain?

Because most figures come from the companies themselves, they measure different things, and they rarely publish full
methods. Independent researchers often have to estimate from the outside.

Things that change the answer:

- **What is included:** just the chips, or also cooling, idle spare capacity, networking, and building the hardware?
- **Which model:** small models use far less than large ones, and "reasoning" models that think step by step
  produce many more words behind the scenes.
- **How long the answer is:** energy rises with the amount of text read and written.
- **Where and when it runs:** the same electricity can be low-carbon in one country or hour and high-carbon in
  another.

So treat any single number as a rough guide. Honest sources say what they included and how sure they are.

## Which everyday habits matter, and which don't?

The habits that matter are those that multiply use: generating lots of images or video, regenerating again and
again, and leaving AI running when you don't need it. Small courtesies and occasional questions matter very little.

**Habits that make a real difference:**

- **Use the right tool.** A normal web search, a calculator, a map or a dictionary is often quicker and uses less.
- **Go easy on images and video.** Generate what you need, not dozens of variations for fun.
- **Refine, don't keep regenerating.** If the first answer isn't right, say what to change rather than pressing
  "try again" over and over.
- **Ask for short answers** when that's all you need. UNESCO and University College London found that shorter
  prompts and responses could cut energy use by more than half.
- **Pick lighter options** where an app offers a "fast", "mini" or "lite" model for simple questions.
- **Switch off AI features you don't use** in apps and devices, where settings allow.

**Habits that matter little:**

- Saying "please" and "thank you". The extra words are tiny next to a long answer.
- Asking an occasional question of a text chatbot. On OpenAI's figure, each one uses about as much energy as a
  high-efficiency light bulb does in a couple of minutes.

The bigger picture is less about individual habits and more about how AI is built into products and services used by
very large numbers of people. That is where pressure for efficiency and honest information makes the most difference. Our
[use AI efficiently](/practice/use-ai-efficiently/) page has more tips.

## How can I judge claims about AI and the environment?

Ask what is being measured, who measured it, and what it is compared with. Both alarming and reassuring claims can
mislead if these are left out.

Questions to ask:

1. **Per what?** Per question, per image, per year, per user? A big total can mean a small amount per use, and the
   reverse.
2. **Who says so?** A company's own figure, an independent study, or a news story repeating an old estimate?
3. **What's included?** Only the computer chips, or the whole data centre, the electricity supply and manufacture?
4. **How old is it?** Figures change quickly. A number from a few years ago may be ten times too high, or too low.
5. **Compared with what?** "Ten times more than a search" means little unless you know how much a search uses.
6. **Is efficiency being confused with less use?** If AI gets cheaper per question, people may simply use more of it.
   This is called the [rebound effect](/glossary/#rebound-effect).

Be wary of words like "green", "clean" or "carbon neutral" without figures behind them. Independent ratings are
starting to appear: the AI Energy Score, launched by Hugging Face and partners, rates AI models from one to five stars
for energy efficiency, a little like appliance labels. Read more in
[energy labels for AI](/articles/energy-labels-for-ai/).

## How can I have a say?

You can influence green AI through what you choose, what you ask for at work or school, and public decisions about
data centres and regulation. Organisations do listen when enough people ask.

- **As a user:** choose services that publish clear energy and water information, and tell companies when they
  don't.
- **At work:** ask whether your employer has a policy on AI use and on choosing suppliers. Point them to our
  [guide for business](/guides/business/).
- **At school or university:** ask how AI and its environmental costs are taught. Teachers can find resources in our
  [schools section](/schools/), and young people can start at [for pupils](/schools/for-pupils/).
- **Locally:** in many countries, new data centres need planning permission. Planning applications and local consultations are
  a chance to ask about electricity, water and waste heat.
- **Nationally:** governments consult on AI, energy and digital policy. Respond to consultations, or write to your
  elected representative, asking for energy and water reporting from AI providers.
- **In your community:** charities and community groups can use AI well too; our
  [guide for charities](/guides/charities/) may help.

## Can AI also help the environment?

Yes, in some uses, and that is part of the picture. AI is used to forecast weather, spot faults on power lines,
plan the use of renewable energy and speed up scientific models, and some of these uses may save more energy than
they consume.

The catch is that such savings are rarely measured, while the costs of everyday AI use are spread across a vast number
of devices and services. Green AI asks for both sides to be counted honestly: what a use of AI costs, and what it
actually saves. It also favours using a simple, well-understood method where one already does the job, and keeping
AI for the problems where it adds something new.

Researchers are working on ways to make AI models much smaller and cheaper to run, often using mathematics rather
than more computing power. That work matters because it makes the same help available for less energy, on more
devices, and to more people.

## What to do next

1. **Next time you use AI**, ask yourself whether a simpler tool would do.
2. **Cut back on image and video generation** you don't really need, and refine instead of regenerating.
3. **Ask for short answers** when you only need a quick reply.
4. **Check one claim** you see about AI and the environment against the six questions above.
5. **Ask your employer, school or council** what their approach to AI and energy is.
6. **Learn more:** start with [what is green AI?](/articles/what-is-green-ai/) and
   [why green AI matters now](/articles/why-green-ai-matters-now/), or explore the [schools section](/schools/) with
   a young person.

## Frequently asked questions

### How much energy does one ChatGPT question use?

OpenAI's chief executive has said an average query uses about 0.34 Wh, about what an oven uses in a little over one second. Epoch AI independently estimated about 0.3 Wh for a typical query. Long or complex requests use more.

### Does AI use a lot of water?

Per text prompt, a small amount: Google reports 0.26 mL for a median Gemini prompt. Totals matter more: researchers estimated that training GPT-3 in Microsoft's US data centres could have evaporated about 700,000 litres of fresh water. Local effects depend on where data centres are.

### Should I stop using AI to help the planet?

Not necessarily. On OpenAI's figure, an occasional text question uses about as much energy as a high-efficiency light bulb does in a couple of minutes. It makes more sense to use AI where it helps, avoid wasteful habits such as generating dozens of images, and push for better information.

### Does saying 'please' and 'thank you' to a chatbot waste energy?

Each extra word adds a tiny amount of processing, but it is very small next to the energy of generating long answers, images or video. It is not a habit worth worrying about.

## Sources

- [Measuring the environmental impact of delivering AI at Google scale](https://arxiv.org/abs/2508.15734), Elsworth et al., Google (2025)
- [The gentle singularity](https://blog.samaltman.com/the-gentle-singularity), Sam Altman (OpenAI) (2025)
- [How much energy does ChatGPT use?](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), Epoch AI (2025)
- [Our contribution to a global environmental standard for AI](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai), Mistral AI (2025)
- [Power Hungry Processing: Watts Driving the Cost of AI Deployment?](https://arxiv.org/abs/2311.16863), Luccioni, Jernite and Strubell, ACM FAccT (2024)
- [Making AI less 'thirsty': uncovering and addressing the secret water footprint of AI models](https://arxiv.org/abs/2304.03271), Li, Yang, Islam and Ren (2023)
- [Energy and AI](https://www.iea.org/reports/energy-and-ai), International Energy Agency (2025)
- [Smarter, smaller, stronger: resource-efficient generative AI and the future of digital transformation](https://www.unesco.org/en/articles/ai-large-language-models-new-report-shows-small-changes-can-reduce-energy-use-90), UNESCO and University College London (2025)
- [Announcing AI Energy Score ratings](https://huggingface.co/blog/sasha/announcing-ai-energy-score), Hugging Face (2025)
