# What is green AI?

> A plain-language definition of green AI: AI for green and greener AI, efficiency and sufficiency, and its footprint in energy, carbon, water and materials.

Source: https://greenai.guide/articles/what-is-green-ai/  
Updated: 2026-11-19

## In short

Green AI means using artificial intelligence in ways that keep its environmental cost in proportion to the value it delivers. It covers two ideas, using AI to help the environment ("AI for green") and reducing the footprint of AI itself ("greener AI"), and two ways of getting there: efficiency (doing each task with less) and sufficiency (doing only the tasks worth doing).

## Key points

- "AI for green" uses AI to help environmental goals; "greener AI" reduces the footprint of AI itself. Green AI needs both.
- The footprint of AI is not only electricity. It includes carbon emissions, water for cooling and power generation, and the materials and waste of hardware.
- Efficiency reduces the cost of each task; sufficiency asks whether the task needs AI, or such a large model, at all.
- Green AI is not "no AI". It is about using AI well, measuring its cost and choosing the right tool for the job.
- Because "AI" covers everything from small classifiers to frontier chatbots, any green AI claim should say which kind of AI it means.

"Green AI" is used to mean several different things, which makes conversations about it confusing. This article
sets out the main ideas and the terms used across this site, so that later pages can build on them.

## What does green AI mean?

**Green AI means using artificial intelligence in ways that keep its environmental cost in proportion to the value
it delivers.** It is about using AI well, not about refusing to use it.

The term was popularised by Schwartz, Dodge, Smith and Etzioni in a 2020 paper in *Communications of the ACM*. They
contrasted "Red AI", research that buys small gains in accuracy with ever larger amounts of computation, with "Green
AI", which reports and rewards efficiency alongside accuracy. Their paper followed Strubell, Ganesh and McCallum's
2019 study, which showed how much energy and carbon developing a large language model could involve.

Since then the meaning has widened. Today green AI usually covers the whole life of an AI system: the hardware it
runs on, the data centre, the model, the software around it, and how people use it.

## What is the difference between "AI for green" and "greener AI"?

**"AI for green" uses AI to help environmental goals; "greener AI" reduces the footprint of AI itself.** Green AI
needs both, and the two can pull in opposite directions.

| | AI for green | Greener AI |
|---|---|---|
| Aim | Use AI to cut emissions or protect nature elsewhere | Cut the energy, carbon, water and materials AI uses |
| Examples | Forecasting renewable output, managing grids, detecting leaks, climate modelling | Smaller models, efficient code, cleaner power, longer hardware life |
| Risk | Claimed benefits that are never measured | Savings that are spent on more use (the rebound effect) |

The International Energy Agency's 2025 *Energy and AI* report covers both sides: it finds that wider use of existing
AI applications could reduce emissions in energy and industry, while warning that AI's own electricity demand is
rising fast and that the net effect depends on how AI is deployed.

A 2026 UK green AI sandpit of mathematicians, computer scientists and industry framed green AI in the same two-sided
way, with mathematics able to contribute to both: better methods for environmental problems, and leaner methods for
AI.

## What are efficiency and sufficiency?

**Efficiency is doing each task with less; sufficiency is doing only the tasks worth doing.**

- **Efficiency** asks: how can this model, prompt or data centre use less energy for the same result? Examples are
  pruning and quantising models, better chips and more efficient cooling.
- **Sufficiency** asks: does this task need AI at all, or a model this large? A calculator, a database query or a
  classical numerical method is often cheaper and more reliable than a large language model.

Efficiency on its own can backfire, because cheaper AI tends to be used more. That is the rebound effect, explained
in [efficiency is not sufficiency](/articles/efficiency-is-not-sufficiency/).

## What does AI's environmental footprint include?

**Four things: energy, carbon, water and materials.** Most public discussion covers only the first two.

### Energy

AI runs on electricity, mostly in data centres. The International Energy Agency estimates that data centres used
around 415 TWh in 2024, about 1.5% of the world's electricity, and expects that to more than double to around 945
TWh by 2030, with AI the most important driver. See [why green AI matters now](/articles/why-green-ai-matters-now/).

Energy is used in two phases:

- **Training**, a large one-off cost to build a model.
- **Inference**, the cost each time the model is used. For widely used models this adds up: Google reported in
  2022 that inference made up about three-fifths of its machine-learning energy use, and training two-fifths.

### Carbon

The carbon footprint depends on how much energy is used **and** how clean the electricity is. The same job can emit
many times more carbon on a coal-heavy grid than on one dominated by wind, hydro or nuclear power. It also includes
**embodied carbon**: the emissions from making chips, servers and buildings. Wu and colleagues at Meta reported in
2022 that embodied carbon was roughly 30% of the footprint of Meta's large-scale machine learning, and argued it was
becoming the dominant source as operational energy gets cleaner.

### Water

Data centres use water for cooling, and power stations use water to generate the electricity. Li, Yang, Islam and Ren
showed in 2023 that this "secret water footprint" of AI is substantial and varies greatly with location and time of
day. Google's 2025 measurement of its Gemini assistant put the median text prompt at about 0.26 millilitres of water,
alongside 0.24 watt-hours of energy.

### Materials

AI hardware needs metals and minerals, and it is replaced quickly. Wang and colleagues estimated in 2024 that
large language models could generate between 1.2 and 5.0 million tonnes of electronic waste in total between 2020
and 2030, depending on how fast their use grows, and that reuse and recycling could cut this substantially. Keeping hardware in service longer, and making software that runs well on older devices, reduces this.

## Why does it matter which "AI" we mean?

**Because "AI" covers systems whose footprints differ by many orders of magnitude.** A small model that sorts
emails, a weather forecasting system and a frontier chatbot are all "AI".

In everyday use, "AI" now mostly means generative AI: chatbots and image and video generators. Much of the concern
about AI's energy use is about these large models. But most machine learning in industry and science is smaller and
more specialised. A good green AI claim says which kind of AI it is about, and over what boundary its costs are
counted. See [measurement is the bottleneck](/articles/measurement-is-the-bottleneck/).

## Is green AI the same as "no AI"?

**No.** Not using AI is sometimes the greenest choice, and green AI includes that option. But green AI is mainly
about using AI responsibly: measuring its cost, choosing the right model for the task and making it more efficient.

Some people prefer other words. "Sustainable AI" is often used to include social and economic as well as
environmental sustainability. Others argue for describing the goal as avoiding *wasteful* computing, which is harder
to dismiss as merely worthy. This site uses "green AI" for the environmental side, and is specific about which
resource (energy, carbon, water or materials) a claim is about.

## What to do next

- Read [why green AI matters now](/articles/why-green-ai-matters-now/) for the scale of AI's energy demand.
- Learn the key terms in the [glossary](/glossary/).
- Find the guide for you: [business](/guides/business/), [researchers](/guides/researchers/),
  [charities](/guides/charities/) or [the public](/guides/public/).
- Start with the practical guide to [using AI efficiently](/practice/use-ai-efficiently/).
- Teaching? See the [For schools](/schools/) section.

## Frequently asked questions

### What is green AI in simple terms?

Green AI is artificial intelligence designed, built and used so that its environmental cost, in energy, carbon, water and materials, is as low as it can reasonably be for the value it delivers.

### Who coined the term green AI?

The term was popularised by Schwartz, Dodge, Smith and Etzioni in a 2020 Communications of the ACM paper, which contrasted 'Red AI', research that buys accuracy with ever more computation, with 'Green AI', which treats efficiency as an evaluation criterion alongside accuracy.

### Is green AI the same as sustainable AI?

The terms overlap and are often used interchangeably. 'Sustainable AI' is sometimes used more broadly, to include social and economic sustainability as well as environmental impact.

### Does green AI mean using less AI?

Sometimes. Using AI only where it adds value is part of green AI, but so is making the AI that is used more efficient, and using AI to cut emissions elsewhere.

## Sources

- [Green AI](https://doi.org/10.1145/3381831), Schwartz, Dodge, Smith and Etzioni, Communications of the ACM (2020)
- [Energy and policy considerations for deep learning in NLP](https://arxiv.org/abs/1906.02243), Strubell, Ganesh and McCallum, ACL (2019)
- [Energy and AI](https://www.iea.org/reports/energy-and-ai), International Energy Agency (2025)
- [Measuring the environmental impact of delivering AI at Google Scale](https://arxiv.org/abs/2508.15734), Elsworth et al., Google (2025)
- [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)
- [E-waste challenges of generative artificial intelligence](https://doi.org/10.1038/s43588-024-00712-6), Wang et al., Nature Computational Science (2024)
- [Sustainable AI: environmental implications, challenges and opportunities](https://proceedings.mlsys.org/paper_files/paper/2022/hash/462211f67c7d858f663355eff93b745e-Abstract.html), Wu et al., MLSys (2022)
- [Good news about the carbon footprint of machine learning training](https://research.google/blog/good-news-about-the-carbon-footprint-of-machine-learning-training/), Google Research (Patterson) (2022)
