Methodology
How we turn an everyday action into a “number of prompts”. Everything is honest and bounded by ranges: orders of magnitude, never false precision.
In short
We start from an action's measured, sourced consumption, divide it by one prompt's consumption, and get its equivalent in AI prompts. Nothing more.
How much does an AI prompt use?
For an average text prompt: one inference of a large language model (LLM) of the GPT-4o class.
We count only inference (usage), not training. Reasoning queries (~15-40 Wh) and image (~6-12 Wh) or video (~940 Wh/5 s) generation cost far more, see the glossary.
The formula
prompts = action's consumption ÷ one prompt's consumption
Three examples, one per metric:
That's the whole mechanic of the game: each added action is converted into its prompt equivalent, then compared to an AI “budget”.
“Virtual” water and CO₂
For food and goods we count the production footprint (“virtual” water and CO₂), not just the tap water at the meal. A burger “costs” mostly through farming (~2,500 L of virtual water), not the glass of water beside it: a toilet flush (direct water) and a steak (virtual water) don't measure the same thing.
The scales: which AI “budget”?
The budget we try to “spend” matches a real AI consumption:
We don't assume everyone uses AI: these budgets rest on actually observed usage (ChatGPT ≈ 2.5 billion prompts a day), not on world population. In the game you aim at one of the annual budgets (You, 100 people, all generative AI); “ChatGPT in 1 day” is shown here as a reference point.
Why ranges instead of one exact figure?
Sources vary (power mix, measurement method, scope). We keep a credible median and show the range. Two telling examples:
- Flight: depending on whether radiative forcing is included (×~1.9 for DEFRA/ADEME; CO₂ only for ICAO) and one-way vs round-trip, the figure varies twofold.
- Beef: from ~60 to ~99 kg CO₂e/kg depending on the farming type.
Showing the range is refusing false precision.
What about training the models?
It matters, and it's far from negligible: training an LLM can use on the order of a gigawatt-hour of electricity, sometimes much more, the equivalent of hundreds of homes for a whole year.
It is, however, a one-off cost amortized over the hundreds of billions of inferences that follow; per prompt, it becomes small again. The site only quantifies inference (the everyday usage each of us controls), but training is very much part of AI's overall footprint: we say so clearly rather than hide it.
FAQ
How much does an AI prompt use?+
For an average text prompt: ~0.3 Wh of electricity, ~0.3 mL of direct cooling water, ~0.2 g of CO₂. We count only inference, not training.
How do you compute an action's “cost in prompts”?+
We divide the action's consumption by a prompt's: prompts = action value ÷ prompt cost. A 50 L shower ≈ 50,000 mL ÷ 0.3 mL ≈ 167,000 prompts.
Why 12,000 prompts per person per year?+
It's a heavy-use assumption (≈ 33 prompts/day). The larger scales don't multiply by world population: they rest on AI's real consumption (ChatGPT ≈ 2.5 billion prompts/day).
What are “virtual” water and CO₂?+
For food and goods we count the production footprint (“virtual” water/CO₂), not just tap water.
Why ranges instead of one exact figure?+
Because sources vary. We keep a credible median and show the spread: honest orders of magnitude, not false precision.
What about training the models?+
Costly but one-off, amortized over hundreds of billions of requests; per prompt, it becomes small. The site quantifies usage (inference), while acknowledging training is part of the overall footprint.
Every value is dated and linked to its source on the Sources page.