Ask whether AI is wrecking the planet and you'll get two confident, opposite answers. One camp says every chatbot reply drinks a bottle of water and boils the atmosphere. The other says a query costs less than a Google search, so relax. Both are selling you something. The honest picture is more interesting — and it hinges on a comparison most of the noise ignores: not AI versus nothing, but AI versus the humans doing the same work.

Let's do this properly, with real numbers and their caveats.

What a single query actually costs

The best-evidenced figure we have comes from Google, which in August 2025 published the most detailed accounting any lab has disclosed. A median text prompt to its Gemini model uses about 0.24 watt-hours of energy, 0.26 millilitres of water — roughly five drops — and emits about 0.03 grams of CO₂. That's seconds of a lightbulb and a splash you wouldn't notice.

Crucially, Google's number is the honest version: it counts idle and standby hardware, cooling, and data-centre overhead, not just the active chip. Measure only the chip and you'd get a flattering 0.10 Wh — which is the trick behind a lot of "AI is basically free" claims. OpenAI's Sam Altman, in June 2025, put an average ChatGPT query at about 0.34 Wh and a third of a millilitre of water — no published method, but reassuringly close to Google's measured figure.

So when you see "one query equals a 500ml bottle of water," bin it. That viral stat traces to a 2024 estimate for a 100-word email under specific conditions, and the researcher behind the original numbers has since said the bottle framing "needs to go away" — the real figure is tens of times smaller. The other zombie stat, "2.9 watt-hours per query," is an order of magnitude too high; it's built on a 2009 Google-search baseline from before modern AI existed.

Your personal AI use, in other words, is a rounding error in your own carbon footprint. One long-haul flight outweighs a lifetime of chatbot queries.

So why all the alarm?

Because "tiny per query" and "huge in total" are both true at once, and they don't cancel out.

The International Energy Agency — the authoritative source here — reports that the world's data centres used about 415 terawatt-hours of electricity in 2024, around 1.5% of global supply, and projects that roughly doubling to ~945 TWh by 2030, with AI the single biggest driver. Cooling those centres consumed an estimated 17 billion gallons of water directly in the US alone in 2023 (per Lawrence Berkeley National Laboratory), heading for two-to-four times that by 2028. And the carbon depends almost entirely on the grid: the same query run in coal-heavy Poland can be about 21 times dirtier than one run on Ontario's nuclear-and-hydro grid.

There's also the half nobody counts: the embodied emissions of manufacturing the chips and pouring the concrete — perhaps 30% of a data centre's lifetime carbon, and a rising share as grids clean up.

And then there's the trap that undoes the comforting per-query maths entirely: the rebound effect, or Jevons paradox. When something gets cheaper and more efficient, we use far more of it — so the total goes up, not down. Microsoft's Satya Nadella said it out loud in January 2025: as AI gets efficient, "we will see its use skyrocket." More efficient per query is exactly why the aggregate footprint keeps climbing.

The comparison that actually matters: AI vs a human

Here's the angle the doom-or-relax debate skips. AI rarely replaces nothing. It replaces a person doing a task. So what does the human cost?

A peer-reviewed 2024 study in Scientific Reports (Tomlinson, Black, Patterson and Torrance) ran exactly this comparison and found, bluntly, that AI emits far less carbon per task than a person doing the same job. Writing a page of text, an AI system produced on the order of 130 to 1,400 times less CO₂ than a human writer. Creating an image, 310 to 2,900 times less than a human illustrator. The wide range is geography: the low end compares against a writer in a low-footprint country like India; the high end against a high-footprint American.

Eye-watering numbers — but here's where most coverage stops and we won't. Look at how they're calculated. The "human" cost isn't the writer's laptop. It's a slice of that person's entire annual carbon footprint — their flights, heating, food, commute — apportioned to the hours the task takes. An American's ~15 tonnes of CO₂ a year works out to roughly 1.7kg per working hour, and a page takes about 0.8 of an hour. That apportioned life-footprint, not the computer, is what dwarfs the AI.

And that exposes the catch the authors themselves are careful about: the human doesn't stop existing. A writer relieved of a page by an AI doesn't vanish in a puff of saved carbon — they go and do something else, possibly something just as emissions-heavy. The study's own arithmetic produces the odd result that a writer who flew less that year would have "lower writing emissions" — which tells you the number is measuring annual lifestyle, not the act of writing. The authors flag it plainly: their analysis doesn't account for job displacement or rebound effects, and they argue for humans working with AI, not being replaced by it.

So the honest reading: per identical task, AI really is dramatically less carbon-intensive than a human doing it from scratch — but "less carbon per task" is not the same as "less carbon for society," because the displaced human keeps emitting and cheaper AI gets used far more.

What to actually take away

• Your individual AI use is environmentally trivial — don't lose sleep over your query count.
• The thing that matters is the build-out: whether the grids powering an exponential wave of new data centres decarbonise fast enough to absorb the demand.
• Per-task, AI is genuinely efficient compared to human labour — a real point in its favour, provided we don't pretend the humans evaporate.
• Be sceptical of any single scary (or reassuring) number. On this topic, the honest answer always comes with a "but here's the catch."

AI's environmental story isn't "catastrophe" or "nothing to see here." It's a fast-growing draw on the world's power and water that is, task for task, remarkably lean — and whose real footprint will be decided less by how often you use it than by how quickly we clean up the grid behind it.