Is ChatGPT bad for the environment?
Not in the way the question usually implies. One query is small — OpenAI's CEO puts it at 0.34 watt-hours, a couple of minutes of an efficient lightbulb. The impact that matters is aggregate: the datacentres, the power and water they draw, and where they are built. Your individual usage is close to a rounding error; the industry's build-out is not, and those are different questions with different answers.
Two honest answers, depending on whose footprint you mean.
Yours. On the only per-query figure OpenAI has published — 0.34 watt-hours, compared by its author to an oven running for a second — a heavy personal chat habit is a trivial share of an individual's energy use. Set against a flight, a winter of heating or a car, it does not register. If you are asking whether to feel guilty about using it, the arithmetic says the guilt is misallocated.
The industry's. This is where the real number lives, and it is not a per-query number at all. Training is substantial on its own: Li et al. estimated 5.4 million litres of water for training GPT-3 in Microsoft's US datacentres, including 700,000 litres consumed on site. Serving hundreds of millions of users daily, continuously, is a larger and permanent load — and it is being met by building datacentres, which consume land, power and water in specific places that have to supply them.
Location is the variable that actually matters. The most useful finding in the academic work is that water efficiency varies substantially by where and when a model runs. A datacentre in a cool, water-rich region is a different proposition from one running evaporative cooling through a hot summer in a stressed basin. That makes this a siting and planning question rather than a personal-consumption one.
What is missing is verification. OpenAI's per-query figures appeared in a blog post without a methodology, scope or measurement period. The peer-reviewed estimates cover an older model and depend on assumptions the authors publish openly. No independent audit of ChatGPT's current environmental cost exists. Anyone offering you a confident single number for 2026 is extrapolating — and we would rather say that than join them.
The arguments, weighed.
Stated at its strongest, and it is not a weak case:
- The load is new and additive. Datacentre demand for AI did not replace an equivalent load; it was added to grids that were already decarbonising too slowly.
- It is concentrated. The burden falls on specific watersheds and specific local grids, whose residents did not choose it and often cannot see the figures.
- Efficiency gains get spent. Cheaper inference has historically meant more inference, not less total consumption — a pattern old enough to have a name.
- Disclosure is voluntary and thin. A parenthesis in a blog post is not a sustainability report, and there is no regulator requiring one.
Also not weak:
- Per query it is genuinely small, and much of the alarming coverage works by multiplying a small number by a large one and reporting the product without context.
- Displacement is real. A question answered in ten seconds may replace a car journey, a printed document or a flight-shaped meeting. Nobody has measured this properly in either direction.
- The comparison class is usually missing. Streaming video, cryptocurrency and ordinary web infrastructure also run on datacentres, and rarely attract the same scrutiny per unit of use.
- The figures being repeated are old. Much of the viral material derives from GPT-3-era estimates, several model generations and efficiency improvements ago.
We will not give you a tonnes-of-CO₂ figure for ChatGPT, a percentage of national electricity, or a "ChatGPT uses as much power as X country" comparison. Those circulate constantly, and within this pass we could not trace them to a primary source we could read and date. Publishing one anyway would be exactly the failure this site exists to avoid.
What we can tell you is which numbers are real, where they came from, what they cover, and how far apart they are — which is on our water page in detail. If a number matters to your argument, it should survive being asked where it came from.
Two things actually change consumption, and both are free. Use a lighter model when the task is simple — energy scales with computation, and ChatGPT's picker exposes reasoning effort directly. Stop re-rolling the same prompt: repeated generations cost repeatedly and, per OpenAI's own account of randomness, tell you nothing about whether the answer is right.
Beyond that, the leverage is not individual. Datacentre siting, water rights and grid planning are decided by regulators and local government, and that is where the version of this question with real consequences gets answered. Choosing to abstain from a chatbot does not participate in that decision.
The comparison nobody can actually make.
We would like to rank assistants by environmental cost and cannot: no vendor publishes audited per-query figures, so any ranking would be invented. These are the honest partial answers instead.
Frequently asked.
Quick follow-ups people search after this question.