Skip to content

AI, decoded

Is using AI bad for the environment?

Probably not at the level of one person's text prompts, though the footprint depends on the workload and what a measurement includes, and some widely shared figures have been wrong by orders of magnitude. The aggregate question is the one that bites: how much power the data-center buildout draws and where that power comes from. Utilities and siting decisions move that far more than how often any one person opens a chatbot. On Chain of Thought, Andy Masley said sending a ChatGPT prompt probably emits as much as running a space heater for about half a second or driving a car about four feet.

· Updated · Chain of Thought

Level 1: How AI works

AI Energy & Data CentersAI Infrastructure

  • 2 episodes
  • 2 guest clips
  • 4 glossary terms
A water bottle beside the words off by ~4,500×, Andy Masley's estimate of the error in one widely cited AI water figure (episode 45), and the line: check the math.

In detail

1. The circulating numbers keep failing basic checks

Andy Masley, who taught physics for seven years, fact-checks AI environmental claims, and the largest error he has found personally is a water figure in the book Empire of AI that is off by a factor of about 4,500 (episode 45 at 7:15). One of the two errors behind it: the author’s source had mixed up liters and cubic meters, and a cubic meter is a thousand liters. The published figure implied that each resident of an 88,000-person community outside Santiago was using about a tenth of a liter of water a day, for everything (7:46). That is a number anyone could have sanity-checked, and for six months nobody did. Same chapter, second problem: the claim that AI would consume half as much water as all of Britain by 2027, where the study being cited puts the figure at roughly 5%.

2. One prompt is a rounding error in your day

Masley’s comparisons, for the conversation you actually have at a party: a single prompt probably emits about what printing a fifth of a page of a book does, or running a space heater for half a second, or driving four feet (39:25). On water, his estimate is roughly one eight-hundred-thousandth of your daily use. Anything you do on a computer carries a small climate cost relative to almost everything else in a day, which is why “I ran a thousand searches today” was never a confession.

3. Overstating the number carries its own environmental cost

Masley’s hedged read of what followed the Santiago coverage: activist pressure, possibly built on the bad figure, appears to have pushed Google toward air cooling instead of water cooling (9:30) for that proposed data center, which has not been built. Air cooling takes considerably more energy, so he suspects the redesign would mean higher emissions in exchange for saving water. His corrected estimate puts the facility at about 25% of the water in one section of Santiago, but about 3% of the municipal supply (9:49). As of September 2026, Google says it has not decided whether to go ahead with the redesigned, air-cooled project. Environmental policy is a series of trade-offs, and you cannot price a trade-off with a number that is wrong by three orders of magnitude.

4. The constraint that binds is the grid

Cisco’s Jeetu Patel names infrastructure as the first hard constraint on AI in episode 44 at 11:39: not enough power, compute, network bandwidth, or data-center capacity to meet demand. His framing is that infrastructure is the input and intelligence is the output, and the goal is generating tokens at the lowest kilowatt and the lowest dollar (15:24). That problem is expensive, contested, and unresolved. It is also a different problem from whether you should feel bad for sending a prompt.

Compare measured work, not a universal prompt number

A short text answer, a long agent run and a generated video are different workloads. One measured case is Google’s August 2025 study, which reports 0.24 Wh for the median Gemini Apps text prompt. That figure covers Google’s own serving infrastructure for that product and period, not every model, workload or provider.

Worked example, with numbers invented for teaching: two workflows each deliver 1,000 successful tasks at the same quality. For this exercise both use an illustrative emissions factor of 0.4 kg CO₂e per kWh.

Metric for 1,000 successful tasksWorkflow AWorkflow BChange
Facility electricity20 kWh12 kWh40% less
Electricity-associated emissions8 kg CO₂e4.8 kg CO₂e40% less
On-site water consumed30 L45 L50% more

Workflow B uses less electricity per success (12 Wh against 20 Wh) but more on-site water, which is a separate input, not derived from the energy figures. The table doesn’t settle which has the smaller total footprint: indirect water, training and other lifecycle impacts are outside it. For a real comparison, measure the same workload and period, count failed attempts in the resources used, and divide by successful tasks.

Why it matters

Individual guilt and grid policy are two different conversations, and running them together costs you both. Bad numbers lose more than an argument. On Masley’s read, one of them may have pushed the cooling design of a proposed data center the wrong way, which is what it costs to reason from a figure nobody checked.

Disclosure: Episode 45 was sponsored by Galileo, the Season 3 presenting sponsor (thanks at 0:00). Episode 44 was sponsored by Galileo while Conor worked there.

AI and the environment: your prompt and the grid are two different conversations, and bad numbers cost both Two panels at very different scales. Your prompt is a dot: Andy Masley's comparisons are that sending one probably emits about as much as printing a fifth of a page of a book, running a space heater for about half a second, or driving about four feet, and his best estimate for water is about one eight-hundred-thousandth of your daily use. The buildout is the large panel: Cisco's Jeetu Patel says there is not enough power, compute, network bandwidth and data-center capacity for AI, and that infrastructure is the input and intelligence the output; utilities and siting move that far more than one person's chatbot use. Underneath, a cascade of what a bad number costs: a water figure off by about 4,500 times, with a liters-versus-cubic-meters mix-up as one of two errors; activist pressure possibly built on it; a data center pushed toward air cooling instead of water cooling, as Masley reads it; and air cooling takes more energy, so he suspects higher emissions, to save water he puts at about 3% of the municipal supply, though about 25% in one section of Santiago. Two conversations, two scales Individual guilt and grid policy are different questions. Running them together costs you both. ONE PROMPT A rounding error in your day Andy Masley's comparisons: it probably emits about as much as a fifth of a printed page, a space heater for ~half a second, or ~four feet of driving. Water, his best estimate: about one eight-hundred-thousandth of your daily use. THE BUILDOUT The constraint that binds is the grid Cisco's Jeetu Patel: not enough power, compute, network bandwidth or data-center capacity to meet demand. "Infrastructure is the input. Intelligence is the output." Utilities and siting move it far more than one person's chatbot use. WHAT ONE BAD NUMBER COST, AS MASLEY READS IT ~4,500× off a water figure; liters vs m³was one of two errors activist pressure,"potentially based onthis misunderstanding" appears to pushGoogle to air coolinginstead of water more energy; hesuspects higheremissions, to save~3% of the city supply
Individual guilt and grid policy are different conversations. Every per-prompt figure is Andy Masley’s hedged estimate from episode 45; the infrastructure constraint is Jeetu Patel’s, episode 44. Download the image

Hear it from the guest

“The number that she gave for this community basically implied that the average person there was consuming about point one liters a day, which is something like a fifth of a single bottle of water”
Andy Masley, Effective Altruism DC EP 45 · around 7:46 · read the transcript
“Sending a prompt probably emits as much as printing a fifth of a single page of a book, or it's like running a space heater for about half a second, or driving your car for about four feet or something like that.”
Andy Masley, Effective Altruism DC EP 45 · around 39:25 · read the transcript

Quotes lightly edited to remove filler words.

From the conversation

This explainer is drawn from these episodes — each carries its full transcript.

Concepts in this explainer

Data CenterAI Energy UseJevons ParadoxTokenization