A teacher’s guide to AI sustainability

Teach AI sustainability as an evidence question, not a rule that every prompt is good or bad. AI services use electricity, cooling water and physical equipment, but a “query” has no fixed footprint. Model size, task, response length, hardware, data-centre overhead and electricity source all matter.

Students should leave able to distinguish energy from carbon and water, trace a claim to its source, explain uncertainty and choose a proportionate tool. Avoid making young people feel responsible for infrastructure decisions controlled by providers and institutions. Personal choices can reduce waste; procurement, product design and public policy shape the larger system. The classroom goal is informed agency: use AI when it supports learning, question unsupported environmental marketing and recognise when ordinary software is enough.

Establish a clear shared model

Begin with a physical description. A prompt travels to a data centre, is converted into tokens and processed by specialised chips. The model reads the input and generates output, usually one token at a time. Networking, memory, cooling and power conversion support that work. What happens when you send an AI message gives a student-friendly sequence.

Then separate three terms on the board. Energy is electricity consumed. Operational carbon depends on how that electricity is generated; broader carbon accounting may include manufacturing and construction. Water can be used in cooling, power generation and chip manufacture, with local effects that depend on scarcity and timing.

The International Energy Agency’s Energy and AI report is a strong system-level source. It covers data centres rather than pretending every watt can be assigned neatly to one classroom question. This distinction helps students hold two ideas together: an ordinary text prompt may be small, while rapidly expanding aggregate demand can still matter.

Teach uncertainty without teaching helplessness

Numbers disagree because their boundaries disagree. One source may measure the accelerator during active computation; another may include idle machines, CPUs, memory, cooling and power losses. A provider average may combine many models and request lengths. A research estimate may model hardware it cannot directly observe.

Give students a four-column source card: product and date; measured or estimated; included boundary; missing information. Ask them to preserve labels such as “median”, “average” and “scenario”. They should never average conflicting figures merely to produce a tidy answer.

Google’s first-party study of Gemini Apps is useful because its method includes active chips, idle capacity, host systems and data-centre overhead. It is also limited to Google’s product, time period and chosen boundary. That combination—useful and limited—is the tone to model.

Treechat’s methodology likewise describes a model-based operational estimate, not a direct meter reading or complete life-cycle claim.

Use discussion prompts that expose trade-offs

The following prompts work in pairs, seminars or a written exit ticket:

• A small model gives a weak answer twice; a larger model answers correctly once. Which is the more sustainable choice, and what evidence is missing?
• Should a school prefer a tool with a low per-prompt estimate or one that publishes lower total annual emissions? Can both metrics matter?
• Is using water for cooling always harmful, or do source, season, scarcity and alternative cooling methods change the judgement?
• If a company funds tree planting, what would you need before calling its AI “carbon neutral”?
• Who has more control over AI’s footprint: a student, a teacher, a school purchaser, a cloud provider or an electricity regulator?
• When does an AI-generated explanation add enough educational value to justify using it instead of search or a textbook?

Do not grade students on reaching one approved moral answer. Grade the quality of the boundary, evidence and reasoning. The AI environmental impact FAQ can serve as a starting reference, not an answer sheet.

Try two honest classroom activities

For a claim audit, give groups a social post or headline about AI water, energy or trees. Students locate the original source and identify the product, year, unit, denominator and exclusions. They rewrite the headline so it remains accurate. This teaches media literacy without needing access to private provider data.

For a task-design exercise, set one learning goal—perhaps explaining photosynthesis to a younger pupil. Groups design three routes: no generative AI, a small text model and a feature-heavy assistant. They predict which steps add value, what data would be sent and which parts likely require more computation. They then assess answer quality and revise their prediction.

Do not present response time, a device battery reading or a provider’s interface icon as a direct energy measurement. If no comparable disclosure exists, “we cannot determine the winner” is a successful finding. UNESCO’s AI competency framework for teachers places sustainability alongside human agency, ethics, foundations and pedagogy; the activity should keep those aims together.

Set a proportionate classroom practice

A workable norm is: use the least complex approved tool that meets the learning purpose. Search for sources. Use a calculator for arithmetic. Use a small model for routine language support. Reserve larger reasoning, image or research systems for tasks where their capability is relevant.

Ask for concise output, trim irrelevant context and avoid generating many disposable variants. But do not turn token counting into a behaviour chart or punish a learner who needs assistive technology. Accessibility, language support and safety can justify computation. Sustainability is one consideration within good pedagogy, not a rationing contest.

Be explicit about privacy and assessment. Follow institutional policy, avoid uploading personal student information, and tell students when AI support is permitted and how it should be acknowledged. UNESCO’s guidance for generative AI in education and research recommends age-appropriate, human-centred use and attention to data protection.

Ask better questions of your institution

Teachers should not be expected to audit a global cloud service alone. Ask procurement teams whether vendors disclose model choices, task-specific energy, data-centre overhead, water boundaries and absolute annual impacts. Ask whether the interface lets users choose a smaller model and whether it makes tool calls visible.

Include students in this enquiry. They can draft questions, compare vendor answers and identify where marketing claims outrun evidence. Our student’s guide to AI environmental impact gives them a parallel checklist, while choosing an AI assistant by energy use explains why model names alone are insufficient.

Keep the answers and review them when contracts, tools or provider disclosures change.

End on agency rather than alarm. A responsible class can use AI thoughtfully, demand clearer disclosure and still acknowledge uncertainty. The lesson is not “never ask”. It is “understand the job, use the right tool and keep the evidence attached to every claim”.