In short
Ask six questions: per what (prompt, image, year?), who measured it, what is included, how old the figure is, what it is compared with, and whether it is about efficiency or total use. Most misleading claims, alarming or reassuring, fail one of these. Published per-prompt figures differ mainly because they count different things.
Steps
- Find the unit the figure refers to, such as per prompt, per image or per year.
- Check who produced it and whether it was measured, estimated or self-reported.
- Look for what the figure includes, from chips only to the whole data centre and hardware manufacture.
- Check the date, because AI energy figures change quickly.
- Ask what it is compared with and whether that comparison is like for like.
- Separate claims about efficiency per use from claims about total use.
Most AI energy claims can be tested with six quick questions. Numbers in headlines are often old, stripped of their units or compared with something that isn't like for like.
The six questions
- Per what? A large annual total can mean a small amount per use, and the reverse. The International Energy Agency estimates data centres used about 415 TWh in 2024, around 1.5% of the world's electricity; that is a total, not a per-prompt figure.
- Who says so? A company's own figure, an independent estimate or a news story repeating either?
- What's included? Google's 2025 paper reports a median Gemini text prompt at 0.24 Wh when it counts host machines, idle capacity and data-centre overhead, and 0.10 Wh counting only active chips. Same prompt, two honest answers.
- How old is it? Epoch AI's estimate of about 0.3 Wh for a typical ChatGPT query is about ten times lower than an older, widely quoted 3 Wh. Google reports its median prompt's energy fell 33-fold in a year.
- Compared with what? "Ten times a web search" means little unless you know the search figure and its date.
- Efficiency or total? Cheaper per use can still mean more in total if use grows: the rebound effect.
Common mistakes
- Comparing figures with different boundaries. Mistral AI's life-cycle assessment, which counts more stages, put a 400-token response at 1.14 g CO₂e and 45 mL of water; that is not directly comparable with a chips-only figure.
- Treating a short question and an image or video as the same thing. They differ by orders of magnitude.
- Dismissing all figures as unknowable. Uncertain is not the same as meaningless; look for a range.
Read why measurement is the bottleneck, and how to spot greenwashing in AI marketing.
Sources
- Measuring the environmental impact of delivering AI at Google scale, Elsworth et al., Google (2025)
- How much energy does ChatGPT use?, Epoch AI (2025)
- Our contribution to a global environmental standard for AI, Mistral AI (2025)
- Energy and AI, International Energy Agency (2025)
Related
How do I spot greenwashing in AI marketing?
Warning signs of greenwashing in AI products and services, from vague 'sustainable AI' labels to offset-based neutrality claims, and the questions to ask.
How do I put my AI energy use in perspective?
How one AI question compares with everyday energy use, why images, video and long documents change the picture, and why the global total still matters.
Measurement is the bottleneck for green AI
Why AI energy figures are hard to trust or compare: self-reported numbers, estimates not meters, different boundaries, and what good looks like.
Honest green claims about AI
How to avoid greenwashing about AI: what you can and can't claim, rebound effects, honest reporting, and the UK, EU and US rules on environmental claims.