Teaching "AI & the Environment" to Hacker mode (ages 11–14)

Hook & Warm-Up

Open with the framing this age band can actually engage with critically:

"AI's environmental paradox: the technology consuming an increasing percentage of global electricity is also one of our best tools for combating climate change. Let's quantify both sides."

Push immediately for skepticism rather than acceptance: "Notice the word 'paradox' — that means both things are true at once, and they don't cancel out neatly. Today we're not deciding if AI is good or bad for the planet. We're looking at the actual numbers on both sides, and at how researchers even try to measure something this complicated." Ask the class: "What would you need to measure to fairly compare AI's environmental cost against its environmental benefit?" (Good answers: the same units on both sides — e.g., tons of CO₂ saved by grid optimization vs. tons of CO₂ emitted by training and running models; a defined time period and scope; who's doing the measuring and whether they have an incentive to make AI look good or bad.)

This is a good moment to name the general skill: comparing a diffuse, large-scale benefit (energy saved across a whole grid, over years) against a concentrated, measurable cost (electricity used in one training run) is genuinely hard to do fairly, and that difficulty is itself part of why the debate hasn't been settled — it isn't just that nobody has bothered to calculate it.

If your class has covered exponential growth in math, this is a good spot to connect it: ask what "consuming an increasing percentage of global electricity" implies mathematically if the trend continues unchecked, versus what it implies if efficiency research (covered later in the lesson) manages to bend that curve. Framing the whole topic as a race between two exponential-ish trends — AI usage growth versus AI efficiency improvement — gives students a mental model to hang the rest of the lesson's facts on, rather than treating each statistic as a disconnected trivia item.

Main Activity

Walk through the lesson's three scenes, but treat each fact as a claim to interrogate rather than accept — this age band should leave the lesson able to ask "compared to what, and measured how?" about any environmental statistic.

Scene 1 — AI's Carbon Cost. "Training GPT-4 used enough electricity to power 1,000 homes for a year." Have students interrogate the claim itself first: OpenAI has never published GPT-4's training energy use, so this specific number is an outside, unverified estimate that has circulated widely online — worth using to convey scale, not as a confirmed fact. Push further: "That's the training cost. What other cost category does the lesson mention that's separate from training?" (Answer: data center operation broadly, and the ongoing electricity of running the model afterward — inference — every time someone actually uses it.) On "Data Centers": "AI data centers use 1-2% of global electricity, and that number is doubling every few years." Flag that this is a recent-estimates snapshot, not a fixed number — figures like this (often IEA-style projections) get revised frequently as AI adoption changes, so students should treat the specific percentage as current-best-guess rather than settled. Have students do quick mental math: if something doubles every few years starting from 1-2%, sketch on the board roughly what a handful of doubling cycles look like (2% → 4% → 8% → 16%) — this is the same exponential-growth intuition used elsewhere in math class, applied to something concrete. On "Water Usage": "Data centers use billions of gallons of water for cooling — Google's AI consumed 5.6 billion gallons in 2023." Correct the framing as you deliver it: that 5.6 billion gallons is Google's total data-center water consumption across all its services — search, Gmail, cloud, AI — not a figure isolated to AI, since Google hasn't disclosed an AI-specific number. Ask: "Why would water use, not just electricity use, matter environmentally?" (Water is a locally scarce resource in many places even when electricity supply isn't a constraint — a data center's water footprint can be a serious local issue even if its carbon footprint is offset elsewhere.) This is worth connecting to the Philippine context directly: several countries, including areas of the Philippines, already deal with seasonal water stress, so a data center's location decision genuinely matters — the same facility could be far less controversial built somewhere water-abundant than somewhere already water-stressed.

Scene 2 — AI Helping the Planet. "AI optimizes power grids, reducing energy waste by up to 15% — saving more energy than AI consumes." Have the class evaluate this claim directly: "Is this saying AI is a net environmental positive overall, or something narrower?" Guide them to see it's a claim about one specific application (grid optimization), not a claim about AI as a whole — a common reasoning error is generalizing from one favorable case to the whole category. On "Smart Farming": "AI drones monitor crop health and reduce pesticide use by up to 90%." Ask what other resources precision agriculture likely also saves (water, fertilizer) and what would need to be true for a farm to actually adopt this (access to the drones/sensors and the AI systems, which is itself a real barrier — smaller and poorer farms often can't afford this without a green solution being genuinely inclusive). On "Ocean Cleanup": "AI identifies ocean plastic from satellite images and helps plan cleanup routes." Ask how such a system would actually be trained (labeled images of plastic vs. water, waves, boats — the same supervised-learning idea used throughout AIQ, applied here to an environmental use case), and what could go wrong (a model trained mostly on one ocean region's lighting and water conditions might perform worse elsewhere — the same data-representation issue AIQ covers in its fairness-focused lessons, showing up again here in an environmental context).

Scene 3 — The Balance. "The big question: does AI's environmental benefit outweigh its energy cost? Researchers debate this actively." Treat "Net Impact" as the scene's real content, not filler — spend real time here. On "Green AI": "Researchers are building smaller, more efficient AI models that use 100x less energy for similar results." Introduce the two named research directions from the quiz explicitly: efficiency metrics like Power Usage Effectiveness (PUE — the ratio of a data center's total power draw to the power actually used by its computing equipment; a PUE close to 1.0 means very little energy is wasted on overhead like cooling), and sparse computation, where a large model activates only a fraction of its parameters for any single input instead of running its full capacity every time — reducing the computation (and therefore energy) needed per use without shrinking what the model is capable of overall. On "Carbon Labels": ask why a standardized disclosure requirement (similar to nutrition labels or vehicle fuel-economy stickers) might change company behavior even without a law forcing efficiency improvements directly (public accountability, competitive pressure, enabling researchers to compare models on a shared basis). Note honestly that, unlike nutrition labels, carbon labeling for AI models is not yet standardized or legally required anywhere — the lesson describes an emerging practice some research papers have voluntarily adopted, not an established regulation.

The Philippine angle — a dual role, not just a footnote. This is the moment to bring the abstract cost/benefit debate home rather than leave it as a general-knowledge exercise. The Philippines sits in the world's most active typhoon corridor and is regularly ranked among the countries most exposed to climate disaster, so the tension in this lesson isn't hypothetical here. On one side, everything in Scene 1 applies directly: any data center serving the Philippines, or any Philippine institution running AI models, draws on the same electricity grid and, in some designs, the same water supply that seasonal water stress already strains in parts of the country. On the other side, PAGASA (the national weather agency) and partner efforts — including AI-based flood forecasting tools like Google's Flood Hub, which now models Philippine river basins — increasingly use machine-learning models trained on historical weather and river-gauge data to extend how much warning families get before a typhoon makes landfall or a river floods. That's not a side benefit; for a country that sees roughly 20 tropical cyclones a year, extra hours of warning save lives on a routine basis. Put the question to the class directly: "Is it possible for the same technology category to be part of the problem — energy and water cost — and part of the solution — disaster warning — for the same country, at the same time?" Push past a comfortable yes/no: the honest answer is that both are true simultaneously, and the harder, more interesting question is how you'd weigh a diffuse, ongoing cost against a concentrated, life-saving benefit that doesn't even show up in the same units.

Discussion

• "The lesson says AI training compute has grown roughly 10x per year, far outpacing Moore's Law (which describes computing power roughly doubling every couple of years). What does it mean for one trend to 'outpace' another, and why might that matter for AI's future energy use specifically?"
• "PUE measures data center efficiency, not AI efficiency directly. Why might a data center have excellent PUE but still consume enormous total energy, or vice versa?"
• "If sparse/mixture-of-experts models cut computation per use, why wouldn't that guarantee lower total energy use across the whole AI industry?" (Bridge to the rebound effect, covered head-on in the Architect script — appropriate here as a preview if a student raises it.)
• "Whose responsibility should it be to measure and disclose an AI model's environmental cost — the company building it, an independent researcher, or a government regulator? What are the tradeoffs of each?"
• "Is it possible for AI's environmental cost and its environmental benefit to both keep growing at the same time? What would that mean for how we should evaluate AI's overall impact?"
• "A data center could be built in a location with cheap land but a water-stressed climate, or a more expensive location with abundant water and renewable power. If you were advising a government on where to allow new data centers, what tradeoffs would you weigh?"
• "PAGASA and international partners increasingly use AI-based forecasting to give Filipino communities more warning before typhoons and floods, while data centers running AI also draw on electricity and, in some designs, water that's already strained in parts of the country. If you were advising a Philippine government agency on whether to invest more in AI-based disaster forecasting, how would you weigh that tradeoff?"

Quiz Walkthrough

Wrap-Up & Extension

Close with: "AI's environmental story doesn't have a tidy ending yet. It's genuinely expensive to train and run at scale, and it's genuinely useful for fighting climate change in specific, measurable ways. The honest, technically literate position is to hold both facts at once and keep asking for real numbers rather than accepting a company's marketing claim — or a critic's alarmist one — at face value."

Extension activity (20-25 minutes): Have students work in pairs to find one real, recent (from the last two or three years) published estimate of an AI model's or an AI company's energy or carbon footprint, using a search engine and looking specifically for a named source (a research paper, a company's own sustainability report, or a reputable technology publication citing a study) rather than an unsourced social media claim. Each pair reports back: what was measured, over what scope (training only, or training plus inference), and what assumptions the source admitted to making. Close by discussing as a class: which pair found the most trustworthy-looking source, and what made it more trustworthy than the others (methodology disclosed, named authors/institution, acknowledged uncertainty, versus a vague headline number with no source)?

If time is short, a faster version works too: skip the independent search and instead give each pair one of the four quiz statistics from this lesson (10x compute growth, PUE, net-impact debate, sparse models) and have them write two follow-up questions a skeptical scientist would ask before fully trusting that statistic — for example, for "10x per year," a good follow-up is "10x measured in what unit — total FLOPs, dollars spent, number of parameters — and over what exact time window?" This keeps the core skill (interrogating a number instead of just repeating it) without requiring open internet research time.