Does asking an AI to do a task really cost more carbon than doing it yourself? This page won't tell you. It'll let
you argue about it with your own numbers.
Every version of this argument, in both directions, comes down to four numbers. One is measured well. One is
measured well but depends on which basis you think is fair. One is measured, moving fast, and runs out past
about sixteen hours. The fourth isn't really a measurement at all — it's a value judgment wearing a unit.
So this page puts all four on sliders, does the two multiplications, and gets out of the way. There's no answer
here it's trying to sell you, mostly because the defensible range of inputs spans about five orders of
magnitude, and anything more confident than a slider would be lying.
Two of the four are impossible to guess cold, so start from a task instead — answer a question, translate a
document, review a contract, build a feature. Picking one fills in the compute and the human hours together,
because how big a job is is the same fact on both sides of the argument. Then move whatever looks wrong.
If you land on a number you like, be suspicious of it. That's more or less the point.
Start from a task
Sets the two log sliders below
Ask & write
Analyze & decide
Make pictures
Build software
A long agentic run — a feature and its tests, from an issue to a diff someone could review.
≈ 500 model calls · 501 Wh of compute against 7.9 h of human work
Your assumptions
AI task
Log scale, five orders of magnitude — an agentic session runs many model calls, not one query.
Hydro or nuclear ≈ 0.10. US average ≈ 0.37. Marginal US ≈ 0.64. Coal-heavy ≈ 0.80.
Displaced human work
Log scale. This is the slider that decides the answer.
Laptop at home ≈ 0.2. Office plus an average commute ≈ 1.5. Fully loaded ≈ 3+.
The answer, per task
At these settings the datacenter emits 1.9% of the CO₂ that the displaced human hours would have.
Human hours avoided
11.91 kg
AI datacenter
226 g
Net CO₂ per task+11.69 kgSaved by using the AI
AI is cleaner by52.8×
Break-even energy26.48 kWhWhere the two sides tie, on this grid
Where these numbers come from
17 presets, 4 defaults, an argument for each, and the sources to
check them against. Where the honest answer is that nobody knows, that's what it says.
The task presets
Two numbers per task, and the energy one is always a count times a per-unit cost. There are three such costs here, two of them measured. A short chat call is 0.24 Wh: Google's median Gemini text prompt, measured full stack — accelerator, host CPU and memory, idle capacity and datacenter overhead, not just the chip. A generated picture is about 7 Wh, which is one measurement times one ratio: Luccioni, Jernite and Strubell measured 2.907 Wh for the costliest image generator they tested, that figure counts accelerator draw alone, and Google's own accelerator-only number runs 2.4× below its full-stack one. The third cost is the judgment call: the presets price a long-context or agentic call at about 1 Wh, roughly four times the median short prompt, because a call carrying a hundred thousand tokens of context, reasoning tokens and a pile of tool output back into the model is not a one-line chat reply. Nobody has published that one. It is the number to attack first.
The human hours are weaker than the energy figures, and they matter more — the hours are what move the answer by orders of magnitude. Two of them are anchored: translation to the industry median of 2,000–3,000 finished words a day, and the software tasks to METR’s time horizons, which are measured in exactly this unit. The rest are my own estimates from doing the work, and yours will differ. Worth noticing what falls out anyway — the coding tasks, by far the most compute-hungry things here, are also the ones the AI comes closest to losing on.
The best-measured number in this calculator is also the smallest one. Google published a production measurement in 2025: the median Gemini text prompt drew 0.24 Wh — about nine seconds of television. The default here is some two thousand times that, because a task is not a prompt. An agentic session runs hundreds of calls, with long contexts, reasoning tokens, retries, and tool output fed back in. Nobody has published a per-completed-task figure for that. The log scale is an admission that reasonable people differ here by three orders of magnitude, not a flourish.
Worth knowing before you do your own arithmetic: the same prompt Google measures at 0.24 Wh measures 0.10 Wh if you count only the TPU and GPU draw. The accelerator is 42% of the energy — host CPU and memory, idle capacity provisioned for reliability, and datacenter overhead are the other 58%. So an estimate built from chip draw alone runs about 2.4× low, and most of the ones you will find quoted are built that way.
The US annual average is 0.37 kg/kWh. The default sits above it deliberately. Average is arguably the wrong basis for a datacenter that adds new load to a grid: what new demand actually causes is the marginal generator, and the EPA puts the US non-baseload rate at 0.64 kg/kWh — nearly double. The default splits the difference, closer to average. That is a judgment call, not a measurement, and it is the first thing to move if you disagree.
Grid intensity is the one input here you can often just look up. If you know which region the compute runs in, use that number instead of this one.
EPA eGRIDEmission rates for all 26 US subregions; eGRID2023rev2 (June 2025) runs ~3% below the figures above
Hours a human would take
And it is the one with the least defensible default. METR measures the 50%-task-completion time horizon — the length of task, measured by how long a human expert takes, that a model finishes half the time. It has roughly doubled every seven months for six years. Rather than bake in a figure that goes stale in a season, this app points you at the live tracker.
METR notes that measurements above 16 hours are unreliable with their current task suite. The top of this slider is therefore past the point where anyone can currently measure — treat it as extrapolation you are doing yourself.
Mostly the commute. The EPA puts an average passenger vehicle at 393 g CO₂e per mile. The Census puts the average US commute at 27 minutes each way, which at driving speeds is something like 15 miles — so a round trip is roughly 12 kg, about 1.5 kg per hour across an eight-hour day. That is where the default comes from, and you can redo it with your own mileage in your head. Slide down toward 0.2 for someone on a laptop at home; slide up past 3 if you want to load in a share of everything else about employing a person.
The genuinely contested part, and the reason this whole page is a toy rather than an answer: the human does not stop existing when the AI does the task. If the displaced hours get spent on something else that emits, the avoided column is fiction. No slider settles that.
Any one of these can move the answer further than all four sliders put together.
Building the hardwareThe carbon spent manufacturing the GPUs and servers, before any of them run a single token.
TrainingAmortized over however many inferences you think is fair — and that denominator is doing a lot of work.
Whether the hours actually disappearThe big one. A human whose task got automated does not stop existing; they do something else, which may emit more or less. If those hours get reallocated rather than freed, the avoided column is fiction. Nobody has a good answer to this, and economists have been arguing the general version of it since 1865.
WaterCooling has a footprint that carbon accounting does not capture at all.
Failure and retriesThe phrase "energy per completed task" quietly assumes you know how many attempts it took. In agentic work that number is frequently not one.
If that list makes the whole exercise look shaky — yes. This is a slider toy, not a study, and it is most useful for
finding out which of your own assumptions the answer actually hangs on.
