Open a chat window, type a question, get an answer in seconds.
Nothing about that moment feels expensive. No spinning meter. No visible machinery. No sense that anything physical just happened.
That feeling of weightlessness is exactly why the question underneath it rarely gets asked:
What actually just occurred, somewhere else, to produce that answer?
Why AI Feels Free
The interface hides almost everything that makes a response possible.
Behind a single prompt sits a chain most users never see — a data centre, specialised processors, networking equipment, cooling systems and electricity pulled from a grid somewhere.
None of that is visible from a chat window.
The cost has been engineered out of sight, not out of existence.
But there is another part of the story that often gets lost: the resource cost of one AI interaction is not a fixed number.
It depends on the model, the workload, the amount of text being processed, the infrastructure running it and what exactly is being measured.
That makes the question more interesting than simply asking whether AI is “good” or “bad” for the environment.
Is Every Casual Query Wasteful?
No.
And this is where some of the public discussion has overcorrected.
Energy use per individual AI task has fallen sharply as models, chips and data-centre infrastructure have become more efficient. The International Energy Agency says energy efficiency per AI task has been improving at an unprecedented rate, while simple text queries typically consume less electricity than running a television over the same period.
Independent research from Epoch AI estimates that a typical GPT-4o text query consumes roughly 0.3 watt-hours under its assumptions. That is an estimate for a particular model and workload, not a universal number for every AI system. Epoch also estimates that very long inputs can increase energy requirements substantially — to around 2.5 Wh for a 10,000-token input and close to 40 Wh for a 100,000-token input, again under its assumptions.
Google has reported a similar direction of travel. In its own measurement of Gemini Apps, Google estimated median energy consumption of 0.24 Wh per text prompt in May 2025 and reported that the energy used by a median prompt had fallen 33-fold over the previous 12 months under its methodology.
So the idea that every ordinary AI question is an environmental disaster simply isn't supported by the available evidence.
Water is more complicated.
AI systems can have water footprints through data-centre cooling and through the generation of the electricity that powers them. But estimates vary substantially depending on the technology, location, electricity mix, cooling system and what parts of the water footprint are counted.
That means a single viral number — whether measured in millilitres or bottles — should not be treated as a universal “water cost of one AI prompt.”
The more defensible conclusion is simpler:
A single casual text prompt is generally a small resource event.
The bigger question is what happens when billions of them become normal.
Where the Resource Problem Gets Bigger
This is where the efficiency story changes.
Not all AI tasks are equal.
The IEA's 2026 analysis says newer energy-intensive applications — including AI reasoning, video generation and agentic tasks — can consume hundreds or even thousands of times more energy per query than simple text generation.
A short question and a huge document are not computationally equivalent either.
And then there is repetition.
Generate the answer.
Regenerate it.
Ask for a different version.
Upload another document.
Ask the model to analyse everything again.
Turn a five-minute task into a 40-minute conversation because the technology makes every additional question feel almost free.
That is where the arithmetic begins to change.
The problem isn't necessarily one more prompt.
It is the fact that friction has almost disappeared.
AI Use Isn't Simply an Education Problem
There is another assumption worth questioning.
When people talk about unnecessary or foolish AI use, it can be tempting to imagine that the problem belongs mainly to people who don't understand technology.
But AI use is common among highly educated people too.
A large global study of 23,218 higher-education students across 109 countries and territories found that students used ChatGPT for activities including brainstorming, summarising texts and finding research articles. The study captured students' early reactions to ChatGPT, collected in early 2024, rather than current worldwide behaviour.
The point is not that educated people use AI irresponsibly.
It is that formal education does not automatically determine how thoughtfully someone uses a new technology.
Someone can be highly educated and still use AI for entertainment, curiosity, reassurance or endless argument.
The same low-friction interface is available to them.
When AI Becomes the Referee in an Argument
People don't only ask AI to explain things.
They can also use it to construct the argument they want to make — or to challenge someone else's.
“Give me evidence that I'm right.”
“Now prove that they're wrong.”
And the conversation continues.
AI starts to look less like a research tool and more like a digital courtroom in which both sides arrive with their own lawyer.
That is partly an observation about how people use conversational systems, rather than a claim that everyone behaves this way.
But research suggests the possibility is worth examining.
A preregistered Nature Human Behaviour study involving 900 participants tested GPT-4 and human opponents in short, controlled debates. Without personal information, GPT-4 performed on par with human opponents; when given basic personal information about the participant, the GPT-4 system outperformed the human comparison in the study. The debates were controlled experiments involving selected propositions, rather than ordinary real-world arguments.
However, an Author Correction published on 3 September 2026 clarified an important limitation: although the Human–AI personalised condition significantly outperformed the Human–Human baseline, the direct difference between personalised and non-personalised GPT-4 did not reach statistical significance (P = 0.07). The broader finding that GPT-4 could be highly persuasive under the study's conditions is therefore more defensible than claiming that personalisation itself has been conclusively shown to provide an additional advantage.
A separate small study involving 39 participants examined people using ChatGPT while debating in an online soccer forum. Across 13 sessions, researchers found participants sometimes prompted ChatGPT for aggressive responses and produced AI-assisted posts containing similar content and logical fallacies. Because the study was small and tied to a specific online debate environment, it is better treated as exploratory evidence rather than a description of AI users generally.
And that brings us to a harder category:
religion, morality and deeply contested social questions.
AI can be useful here.
It can explain what different religious traditions believe.
It can compare interpretations.
It can summarise historical sources.
It can show where different schools of thought disagree.
It can help someone formulate a question for a scholar, theologian, historian or other expert.
But that is different from settling the disagreement.
A 2025 AAAI/ACM study, Sacred or Synthetic?, by Farah Atif, Nursultan Askarbekuly, Kareem Darwish and Monojit Choudhury, evaluated large language models on Islamic jurisprudence across the four major Sunni schools of thought in both Arabic and English. The researchers found significant variation across models, languages and legal schools, and the study reported lower performance in Arabic. They also examined whether models could recognise when they should abstain rather than confidently give an incorrect answer.
Religious questions are not always simply requests for factual information.
They can involve different schools of interpretation, different authorities, historical context and questions of faith and values.
AI can organise those differences.
It cannot automatically make them disappear.
And perhaps the most important danger is not that AI always gives the wrong answer.
It is that a fluent answer can make uncertainty feel like certainty.
Should People Feel Guilty for Enjoying AI?
Not necessarily.
Curiosity is legitimate.
Entertainment is legitimate.
Experimentation is legitimate.
People have always spent resources on things that were not strictly necessary. Not everything in human life has to pass a productivity test.
A person asking an AI to write a silly poem, explain a strange question or generate an amusing image isn't automatically doing something environmentally irresponsible.
The more useful distinction is between individual impact and system-wide scale.
One prompt may be tiny.
Millions of prompts are infrastructure.
Billions of prompts become an industry.
And increasingly sophisticated prompts can require vastly different amounts of computation.
For users in Pakistan, that infrastructure question can feel less abstract when electricity supply, generation capacity and energy costs already shape everyday life. But the AI question is global: the same infrastructure has to support users across countries and continents.
That is why the environmental question shouldn't be reduced to:
“Should I feel guilty every time I use AI?”
A better question is:
“What kind of AI use are we normalising, and what happens when billions of people do it continuously?”
The Contradiction at the Centre of All This
Here is the tension the IEA's latest data makes difficult to ignore:
AI is becoming more efficient and more resource-hungry at the same time.
Energy use per task is falling.
At the same time, demand for the infrastructure running AI is growing. The IEA reports that global data-centre electricity demand grew 17% in 2025, while electricity consumption from AI-focused data centres grew 50%. Its central projection sees overall data-centre electricity consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030.
This does not prove that efficiency gains themselves caused the growth.
The IEA identifies several forces operating simultaneously: improving efficiency, rapidly increasing adoption and new AI capabilities that enable more energy-intensive uses.
That is where the rebound effect becomes an interesting possibility.
When something becomes cheaper and easier to use, people may use more of it.
AI may be creating a particularly strange version of the same phenomenon.
The scarce resource isn't only electricity.
It is also human attention.
When intelligence becomes almost frictionless, we may begin asking questions we previously wouldn't have bothered asking, generating things we previously wouldn't have generated, debating things we previously would have left alone — simply because the cost of asking has become so close to zero.
That may be the most interesting question surrounding AI.
Not:
“Does one prompt waste too much energy?”
But:
“What happens when billions of people stop feeling any reason to ask whether another prompt is worth asking?”
Primary Sources & Academic References
- International Energy Agency (2026) — Key Questions on Energy and AI
- Epoch AI — How Much Energy Does ChatGPT Use?
- Google (2025) — Measuring the Environmental Impact of Delivering AI at Google Scale
- Ravšelj et al. (2025) — Higher education students' perceptions of ChatGPT: A global study of early reactions, PLOS ONE (Correction, May 2026)
- Salvi et al. (2025) — On the conversational persuasiveness of GPT-4, Nature Human Behaviour (Author Correction, 3 September 2026)
- Zeng et al. (2025) — “Ronaldo's a poser!”: How the Use of Generative AI Shapes Debates in Online Forums, CHI 2025
- Atif et al. (2025) — Sacred or Synthetic? Evaluating LLM Reliability and Abstention for Religious Questions, AAAI/ACM AIES
Do you think about the resource cost behind an AI response before you hit send — or does AI genuinely feel free? And when you use it in an argument, are you looking for information, or proof that you're right? Share your take in the comments.
Comments
Post a Comment