# Use AI efficiently: everyday habits

> Everyday habits that cut the energy of using AI: clear prompts, fewer regeneration loops, care with images and video, caching and turning off defaults.

Source: https://greenai.guide/practice/use-ai-efficiently/  
Updated: 2026-11-19

## In short

Most of the energy of everyday AI use comes from how much it is asked to read and write, how often it is re-run, and which kind of output is requested. Clear, short prompts, fewer regeneration loops, restraint with image and video generation, and switching off AI features you don't use make a real difference without giving up the benefits.

## Key points

- Energy rises with the length of what AI reads and writes; UNESCO and UCL found shorter prompts and responses can cut energy use by over 50%.
- Long inputs are expensive: Epoch AI estimates a 100,000-token input raises a typical query from about 0.3 Wh to almost 40 Wh.
- Image generation was the most energy-intensive AI task in a 2024 study: a mean of 2.9 kWh per 1,000 images, against 0.002 kWh per 1,000 text classifications.
- Regenerating again and again is the hidden cost: plan the request, then refine rather than restart.
- Turn off AI features you don't use, and reuse good outputs instead of regenerating them.

## What uses the energy when I use AI?

**Three things matter most: how much text the AI reads and writes, how many times you run it, and what kind of
output you ask for.** The model's size and the data centre matter too, but those are mostly the provider's choice;
your habits are yours.

Language models process text in [tokens](/glossary/#token), chunks of a few characters. Every token you send and
every token returned is computed. Epoch AI estimates that a typical ChatGPT query uses roughly 0.3 Wh, but a
10,000-token input (about a short paper) raises that to around 2.5 Wh, and a 100,000-token input to almost 40 Wh.
A 2025 UNESCO and UCL report found that shorter, more concise prompts and responses can cut energy use by over 50%.

None of this means a single question is a large energy cost. It means that habits repeated many times, across many
people, add up.

## How do I write prompts that waste less?

**Say clearly what you want, give only the context needed, and ask for the length you need.** A well-planned
request usually gets a usable answer first time, which is the biggest saving of all.

- **Plan before you type.** Decide the purpose, audience, format and length.
- **Give context once, and only what's relevant.** Paste the paragraph that matters, not the whole report.
- **Ask for the length you need:** "three bullet points", "under 150 words", "just the corrected sentence".
- **Ask for one thing at a time** when tasks are distinct; combine related steps when they belong together.
- **Turn off extended reasoning or "deep research" modes** when a quick answer will do. They generate far more
  hidden text.
- **Skip the pleasantries if you like.** A 2026 study measured the energy of short polite messages such as "thank
  you" to show that every extra exchange has a cost at scale. It is a small saving per message, but a free one.

## How do I avoid regeneration loops?

**Refine the answer you have rather than starting again, and stop when it is good enough.** Repeated regeneration
is one of the largest and least visible sources of waste.

Participants at a 2026 UK green AI sandpit of mathematicians, computer scientists and industry described people
running image or video generation 25 times and still not getting what they wanted. Each attempt costs the full
price.

To break the loop:

1. **Write a better brief, not more attempts.** If three tries miss, the request is the problem.
2. **Edit, don't regenerate.** Ask for a specific change ("make the second paragraph shorter") rather than a whole
   new answer.
3. **Use the right tool.** A spell-checker, a template or a search engine may be quicker and lighter.
4. **Set a stopping rule:** for example, at most three attempts, then rethink or do it by hand.
5. **Accept "good enough" for low-stakes work.**

## Why do images and video cost more?

**Generating images, and especially video, needs far more computation than most text tasks.** Choose them when
they add real value.

In the 2024 study by Luccioni, Jernite and Strubell, image generation was the most energy-intensive task measured,
with a mean of 2.9 kWh per 1,000 images, against 0.002 kWh per 1,000 text classifications. For scale, the authors
note that charging an average smartphone takes 0.022 kWh. The least efficient image model they tested used 11.49
kWh per 1,000 images. Video generation produces many frames per clip, so expect it to cost more again, though
comparable public figures are scarce.

Habits that help:

- Sketch or describe first; generate only when the idea is clear.
- Generate at the resolution and length you need, then upscale only the one you keep.
- Reuse images you've already made, or use existing photos and stock libraries.
- Don't generate decorative images by default for every document or post.

## What can teams do with caching and reuse?

**Store and reuse good outputs, and let systems reuse what they have already processed.** Generating the same thing
twice is pure waste.

For everyone:

- Keep a shared library of good prompts, templates and approved outputs (standard replies, summaries, images).
- Search the library before generating something new.

For people building AI into tools:

- **Cache responses** to identical requests so the model is not called again.
- **Use prompt caching** for fixed instructions and shared context. Major providers reuse already-processed input
  for repeated prompt prefixes; OpenAI says this can cut time to first token by up to 80% on long prompts.
- **Don't resend the whole conversation** each turn: cap history, summarise older turns and retrieve only what is
  relevant.
- **Give agents a stop condition** and a cap on steps, so they don't loop. See
  [agentic AI and token waste](/articles/agentic-ai-and-token-waste/).

## Which defaults should I turn off?

**Switch off AI features you don't use, and choose tools that let you opt in rather than out.** AI is increasingly
added to search, email, office software and phones, often running whether or not you wanted it.

The Green Web Foundation suggests organisations ask vendors whether bundled AI features can be turned off. At the
sandpit, participants argued that AI-generated search summaries and similar features should be off by default, so
that AI runs only where it is wanted.

Look at:

- AI summaries in search engines and browsers
- automatic AI drafting, rewriting or meeting summaries in office and email tools
- AI photo enhancement and assistants on phones
- "always-on" assistants and agents in workplace tools

If you manage tools for an organisation, set the default to off and let people turn features on when they need
them.

## Checklist

- [ ] I know what I want before I prompt
- [ ] I paste only the relevant part, not the whole document
- [ ] I ask for the length and format I need
- [ ] I refine rather than regenerate, with a stopping rule
- [ ] I use images and video only when they add value
- [ ] I reuse good outputs from a shared library
- [ ] I've turned off AI features I don't use
- [ ] I start a new chat for a new topic

## What to do next

1. Pick one habit from the checklist and try it for a week.
2. Turn off one AI default you never use, at work or on your phone.
3. Start a shared prompt and output library for your team.
4. If you build AI features, add caching and output-length limits, and [measure the energy](/practice/measure-ai-energy/).
5. Check you're using a right-sized tool; see [choose the right model](/practice/choose-the-right-model/).
6. Share the [public guide](/guides/public/) with friends, or the [lesson ideas](/schools/lesson-ideas/) with a
   school.

## Frequently asked questions

### Does saying please and thank you to AI waste energy?

Every message costs some energy, and a 2026 study measured the cost of short polite messages to make that visible at scale. The bigger savings come from avoiding long inputs, long outputs and repeated regeneration.

### Is generating images worse than generating text?

Usually, yes. In a 2024 study, image generation was the most energy-intensive task tested, with a mean of 2.9 kWh per 1,000 images against 0.002 kWh per 1,000 text classifications.

### Should I stop using AI to save energy?

Not necessarily. Use it where it helps, choose the lightest tool that does the job, and avoid waste. That is the point of green AI.

### Does a long chat cost more than a new one?

In many chat tools the whole conversation is processed again with each turn, so very long threads can grow expensive. Start a fresh chat for a new topic and paste in only what is needed.

## Sources

- [AI large language models: new report shows small changes can reduce energy use by 90%](https://www.unesco.org/en/articles/ai-large-language-models-new-report-shows-small-changes-can-reduce-energy-use-90), UNESCO (2025)
- [How much energy does ChatGPT use?](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), Epoch 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 (arXiv 2311.16863) (2024)
- [Small Talk, Big Impact: The Energy Cost of Thanking AI](https://arxiv.org/abs/2601.22357), Delavande, Pierrard and Luccioni (arXiv 2601.22357) (2026)
- [Prompt Caching 101](https://developers.openai.com/cookbook/examples/prompt_caching101), OpenAI Cookbook (2024)
- [Thinking about using AI? Here's what you can and (probably) can't change about its environmental impact](https://www.thegreenwebfoundation.org/publications/report-ai-environmental-impact/), Green Web Foundation (2024)
- [Green Software Patterns: optimise agent orchestration to reduce model calls](https://github.com/Green-Software-Foundation/patterns/blob/main/docs/development/optimize-agent-orchestration-reduce-model-calls.md), Green Software Foundation (2025)
