In short
Estimate the AI's own footprint (training or fine-tuning, plus running it for its expected life) and set it against the benefit, measured over the same period. Compare it with the simplest non-AI option, not with doing nothing. If the project is meant to save energy or emissions, the saving has to clearly exceed the AI's own cost, or it is a net loss.
Steps
- Write down the benefit you expect and how you will measure it.
- Estimate the AI's footprint over its expected life, including training, inference and retries.
- Compare it with the simplest non-AI method that could do the job.
- Decide in advance what result would make you stop or scale back.
- Re-check with measured figures once the system is running.
Count the AI's own footprint and set it against the benefit over the same period. Projects that use AI to optimise something, including "AI for green" projects, rarely count the energy of the AI itself, and without that number a project can be net negative.
How do I estimate the AI's side?
- Training or fine-tuning, if you do any.
- Inference: requests per day × energy per request × days in service. Epoch AI's estimate of about 0.3 Wh for a typical ChatGPT query is a starting point for text, but long inputs raise it sharply, so use your own token counts; see estimate the energy of AI API calls.
- Retries and failed runs, which are easy to forget.
What should I compare it with?
The simplest non-AI option, not doing nothing. AI is not free even when it saves time: a 2025 study comparing manual and AI-assisted coding on 12 programming tasks found the AI-assisted approach averaged about 33 times the carbon footprint of the manual approach. That does not mean never use AI; it means check that it pays back.
GreenCode, an ITEA research project, publishes a climate case that sets its own AI footprint against the energy saving it aims to make, which is a useful model to copy.
What makes the biggest difference
Deciding the stop condition before you start. Write down the result that would make you scale back, so the decision is not made by momentum.
Common mistakes
- Counting the benefit over ten years and the AI cost over one month.
- Ignoring growth in use once the tool is popular.
- Treating uncertainty as zero. Give a range.
Read efficiency is not sufficiency for the wider picture.
Sources
- Green AI and energy-aware computing, GreenCode (2026)
- Comparative analysis of carbon footprint in manual vs. LLM-assisted code development, Cheung, Kaul, Jahangirova, Mousavi and Zie, ResponsibleSE 2025 (ACM) (2025)
- How much energy does ChatGPT use?, Epoch AI (2025)
Related
How do I decide whether I need AI for a task at all?
A quick test for whether a task needs generative AI or whether search, a calculator, a template or a simple rule would do it faster and with less energy.
How do I stop AI efficiency gains being cancelled out by more use?
How to keep AI efficiency savings from being swallowed by growing use: track totals, cap or budget use, retire freed capacity and ask if it is needed.
How do I estimate the energy of my AI API calls?
You can't meter a hosted AI API, but you can estimate energy per call from token counts and model assumptions with EcoLogits, and log it next to cost.
Efficiency is not sufficiency: the rebound effect in AI
Why making AI more efficient does not guarantee it uses less energy overall, what the rebound or Jevons effect is, and what helps keep savings real.