A student’s guide to AI environmental impact

AI has an environmental impact because the computers behind it use electricity, need cooling and must be manufactured. That impact is real, but there is no reliable fixed footprint for “one AI question”. A short text response from a small model is a different job from generating images, analysing a book or running a research agent.

The best student response is not to panic or pretend the impact is zero. Use the simplest suitable tool, keep requests focused, check environmental numbers before repeating them and ask providers for evidence. Your individual prompt is only one part of the picture: technology companies choose the model, chips, data centre, cooling system and electricity supply that determine most of the footprint.

What happens when you press send

Your message travels over a network to a data centre. The service formats the conversation, turns text into tokens and routes the request to a model. Specialised processors first read the input, then generate an answer token by token. Cooling, networking and power conversion support those processors.

More work usually means more energy. Long input, long output, large models, images, video, reasoning modes and tools can all increase computation. That does not make every advanced feature “bad”; it means the resource should be proportionate to the benefit.

For a closer technical walkthrough, read what happens when you send an AI message. Remember that a message is an interface event, not a scientific unit. Two messages can trigger very different work behind the screen.

The International Energy Agency’s Energy and AI analysis also warns against confusing one request with the system. Data-centre electricity demand depends on both efficiency and the huge number of services being used.

Energy, carbon and water are different

Energy is the electricity used. Carbon refers to greenhouse-gas emissions associated with that electricity and with equipment or construction, depending on the boundary. The same amount of energy can cause different operational emissions on different grids or at different times.

Water may be consumed directly in evaporative cooling and indirectly in electricity generation or manufacturing. Its local effect depends on the catchment, season and source. A memorable bottle comparison can conceal all of that. Google, for example, reported 0.26 ml for a median Gemini Apps text prompt, but explicitly tied the figure to its own 2025 workload and fleet methodology. It is not a universal AI constant.

Do not add energy, carbon and water into one imaginary “eco score” unless the method explains how. They answer different questions. Our AI environmental impact FAQ is a useful place to check common claims and boundaries.

How to check a dramatic claim

Start with the original source, not a repost. Ask four questions: What product and date does it cover? Is it measured or modelled? What is included? Is the result an average, median, scenario or maximum?

Watch for stale comparisons. AI hardware and services change quickly, while popular articles can keep recycling old figures. Watch for denominator tricks too: “per query”, “per response” and “per task” may not describe the same thing. A system that searches several pages and calls a model repeatedly can look like one action to the user.

Conflicting numbers are not always evidence that one side lied. Studies may use different models, locations, cooling boundaries and assumptions. Keep the labels attached rather than averaging them into a number no source found.

Treechat follows this rule in its methodology: the receipt is a model-based operational estimate, not a direct power or water reading and not a full life-cycle assessment.

Use AI when it adds something

For finding a known webpage, use ordinary search. For arithmetic, use a calculator or spreadsheet. For spelling, a local editor may do the job. Use AI when conversation, synthesis, explanation or drafting materially helps.

When you do use it, ask a clear question and specify the length you need. Remove irrelevant pasted material. Continue a useful conversation rather than regenerating from scratch, but start fresh when a huge old context is no longer relevant. Use a small model for routine work and a larger one when the task genuinely requires it. Choosing an AI assistant by energy use offers a practical checklist.

None of this means accepting poor work to save energy. An incorrect small-model answer that causes several retries may be worse than one successful request to a stronger model. Quality, privacy, accessibility and academic rules still matter.

If you are unsure, compare the completed task rather than the number of clicks. A useful result is the fairest denominator.

Keep academic judgement in charge

Check your school or university’s AI policy before submitting work. Cite sources you actually read. Never treat a chatbot’s confident wording as evidence, and do not upload classmates’ information, unpublished research or personal data without permission.

AI can help you form questions, compare explanations, practise a language or critique a draft. It should not replace the learning the assignment is designed to assess. UNESCO’s guidance for generative AI in education recommends a human-centred, age-appropriate approach that protects agency and privacy.

That principle also applies to environmental choices. Students should not carry all responsibility for infrastructure they cannot see. Ask your institution which tools it buys, whether vendors publish energy and water data, and whether a smaller approved model could meet common needs.

Ask who can change each part of the system. That turns a vague feeling of responsibility into a practical conversation about control.

Turn concern into a useful project

Try an evidence audit. Pick three claims about AI energy or water, find their original sources and record the model, date, unit and boundary. Highlight what each source cannot tell you. This exercise is more valuable than making a poster around a single shocking statistic.

You could also compare tools on the same modest task. Record answer usefulness, input and output length, whether web search or other tools ran, and any provider energy disclosure. Do not pretend latency or laptop battery is a complete measurement; use the exercise to reveal what remains hidden.

Finally, write one question for the provider or your institution. Good options include: “Does this figure include idle capacity and cooling?”, “Can users choose a smaller model?” or “Is water reported for the local catchment?” Our companion teacher’s guide to AI sustainability has discussion prompts for a class. Thoughtful scepticism—not guilt—is the skill worth practising.