Congratulations, You Saved the Planet by Not Asking ChatGPT

01 / THE PREMISE

Everyone hates AI. Everyone also uses it.

The obvious expectation is that people who are deeply concerned about AI would avoid using it. In reality, adoption keeps rising while public trust remains low.

A 2026 Pew survey found that 49% of American adults had used an AI chatbot, up from 33% in 2024. Among adults aged 18–29, usage reached 66%. At the same time, 63% said AI was advancing too quickly, and only 16% expected its long-term effect on society to be positive.¹

This is not automatically contradictory: you can use something while distrusting it, and you can believe a technology is useful while opposing the companies, labor practices, copyright rules, politics, or social consequences surrounding it. I believe those are valid discussions.

However, if someone says using an LLM is environmentally unacceptable, continues using it, and then blames other users for aggregate demand, they are not standing outside the system they criticize.

Total inference demand = users × queries per user × computation per query.

Of course, adoption is not driven only by individual chatbot users. AI is increasingly embedded in search engines, office software (I am looking at you, Google Docs), customer service, advertising, social media, and corporate workflows, but aggregate usage does not appear from nowhere; it is composed of individual, corporate, and automatic uses, including the use of people who say they dislike it.

02 / THE SCALE

The global picture is not small.

Obviously, a small footprint per query does not mean the AI industry has a small total footprint.

Data centers consumed roughly 415–460 TWh of electricity worldwide in 2024. The International Energy Agency projects approximately 945 TWh by 2030, just under 3% of global electricity demand and more than double today’s level.²

AI is the largest driver of the expected growth, although those facilities also serve cloud computing, streaming, storage, cryptocurrency, websites, and the rest of the internet. Hardware associated with AI workloads is projected to grow around 30% annually through 2030.

Renewables may meet almost half of the additional electricity demand, but natural gas and coal are still expected to supply more than 40% of the increase. Data-centre electricity emissions are projected to peak at roughly 320 million tonnes of CO₂ in 2030.³

They should also be regulated. But this standard should not be invented only for AI. Fashion companies, airlines, meat producers, oil companies, electronics manufacturers, delivery platforms, streaming services, and social-media companies also benefit from weak reporting, selective accounting, greenwashing, and insufficient regulation.

The answer is consistent environmental governance, not treating one new industry as uniquely sinful while accepting opacity everywhere else.

Corporate overbuilding is part of the story.

It is easy to imagine data-center construction as a direct response to the number of questions ordinary people ask, but the relationship is not that simple. Technology companies build ahead of demonstrated demand because they fear becoming compute-constrained.

Infrastructure decisions also reflect model training, enterprise contracts, cloud ambitions, competition, investor expectations, and the strategic fear of not owning enough capacity. Meta’s reported talks to lease as much as $10 billion in compute capacity to Anthropic illustrate how infrastructure can be accumulated for one strategy and commercialized for another.⁴

This does not make users irrelevant, because demand still matters; it means a new data center is not necessarily being constructed because you asked ChatGPT where to place a comma.

03 / THE BOUNDARY

I am sorry for disappointing you, but there is no universal “AI query.”

A short answer, a long reasoning task, an image, and an AI-generated video do not consume the same amount of energy. Model, hardware, grid, response length, batching, utilization, and cooling all matter.

I know, not the most reliable source, but Sam Altman stated that an average ChatGPT query uses about 0.34 Wh of electricity and 0.32 mL of water. OpenAI did not publish a complete methodology, so these numbers are not ground truth.⁵

An independent bottom-up analysis nevertheless estimated the same 0.34 Wh median for an ordinary frontier-model query, with an interquartile range of 0.18–0.67 Wh. A longer test-time-compute query reached 4.32 Wh.⁶ Google reported 0.24 Wh and roughly 0.26 mL of water for its median Gemini Apps text prompt.⁷

That last figure is much larger, which is exactly why ordinary text, long reasoning, agents, image generation, and video should not be collapsed into one moral unit called “AI use.”

Training is missing.

Per-query estimates describe inference: what happens when an already-trained model generates an answer. They do not fully include pretraining, post-training, failed experiments, architecture searches, model updates, chips, servers, networking equipment, or data-center construction.

These costs matter. However, there is no transparent public dataset that lets us defensibly divide the full development cost of proprietary frontier models by their lifetime queries, so the honest conclusion is not that training is negligible; it is that its amortized contribution per request is unknown.

04 / THE MIRROR

Now compare it with the rest of your life.

The point is not to prove AI harmless. It is to show the scale of familiar activities that rarely receive the same personal moral scrutiny.

You can buy fewer clothes and keep them longer, walk or take rail where possible, share a car or combine errands, replace some beef meals, reduce autoplay, fly economy, travel less often, and stay longer. None of this requires pretending that pleasure itself is an environmental crime.

“Just Google it” is not a moral philosophy.

People opposed to LLM use often say, “Just search Google,” “Just write it yourself,” or “Just do the task without AI.” Fine, but now apply that logic consistently.

Most of these things are not strictly required for survival. We do them because they save time, create comfort, entertain us, express identity, reduce effort, or make life more enjoyable.

That does not make them immoral, but it does mean “you could have done without it” is not, by itself, a serious environmental framework.

Instead, people often treat old habits as natural and new habits as choices. Nobody appears personally offended by the environmental footprint of paperback books; introduce a Kindle, however, and suddenly everyone remembers mining, batteries, electricity, and electronic waste. Familiar consumption becomes invisible, while technological novelty remains morally conspicuous.

05 / THE PLEASURE PRINCIPLE

This is not an argument against joy.

I am not arguing that everyone should stop drinking coffee, traveling, buying clothes, eating enjoyable food, watching films, or using technology. Environmental ethics should not require turning life into a punishment.

If coffee brings you joy, and caffeine, drink your coffee.

If traveling to Japan is meaningful to you, I am not going to pretend that staying home forever is a real alternative.

If a meal connects you to your family or culture, or you simply enjoy it a lot, its value cannot be reduced to one carbon number.

If an LLM helps you learn, work, create, communicate, be productive, or manage a disability, that benefit is also real.

Environmental impact is one consideration among many, not the only value in human life. My point is not to attack your desires; my point is the selective moral superiority involved in accepting the environmental cost of your own pleasures while treating someone else’s useful or enjoyable activity as uniquely frivolous.

And then there is having a child.

I could not finish this article without mentioning this, right? IYKYK. One influential 2017 study estimated that having one fewer child in a developed country could reduce attributed emissions by 58.6 tonnes of CO₂e per year.¹⁶

This number requires a major disclaimer: it assigns parents a fraction of their descendants’ future emissions, so it is a long-term accounting method, not the emissions released by the biological act of having a baby.

I am not arguing that people should not have children. Okay, maybe I kind of do that; at least, I prefer you doing that. But reproductive choices are deeply personal, and reducing a human life to a carbon calculation would be grotesque. However, apparently, I am kind of grotesque.

If you are planning to create an entirely new consumer of food, housing, transport, clothing, electronics, and energy, perhaps do not build your environmental identity around scolding someone for asking a chatbot three questions.