AIQ
AIQ
Part of the AI & the Environment lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · 💻 Hacker (11–14)
Architect mode has no mascot and expects a straight, adult framing — treat this like the opening of a real briefing, not a game introduction. Open with:
"Training one large AI model can produce as much CO₂ as five cars over their lifetime. Let's explore AI's environmental cost — and how AI fights climate change."
Immediately push past the headline number: "That figure comes from a specific, often-cited estimate for a specific class of large model — it's an order-of-magnitude illustration, not a universal constant. Today's goal is to understand the actual cost categories well enough that you could evaluate a real claim like this one yourself, and to look seriously at where AI's climate benefits are real versus overstated." This sets the tone for the whole session: skepticism toward round numbers, precision about what's actually being measured, and comfort with an unresolved answer.
Frame the stakes for this age group directly, since several will go on to study or work in computing: this is a live, contested area of the field they may enter. Companies building large models face real pressure — regulatory, reputational, and increasingly financial (data center electricity costs are a genuine line item) — to measure and reduce this footprint, and "I don't know how to think about this" is not an acceptable answer for someone entering the field. Ask: "If you were hired as an engineer at an AI company tomorrow and your manager asked you to help reduce the environmental cost of your product, where would you even start looking?" Collect a few guesses before moving into the activity — you'll return to this question in the wrap-up.
It's worth being explicit about what this lesson is and isn't asking of students. It is not asking them to arrive at a verdict — "AI is net-positive" or "AI is net-negative" for the climate — because no such verdict currently has broad scientific consensus behind it. It is asking them to leave able to evaluate a specific claim (a company's efficiency announcement, a critic's alarming headline, a journalist's summary of a study) with the right vocabulary and the right questions, which is a more durable and more honest skill than memorizing whichever number happens to be current today.
Work through the lesson's three scenes as a technical case study, treating the specific numbers as jumping-off points for the underlying concepts rather than facts to memorize.
Scene 1 — AI's Carbon Cost. "Training GPT-4 used enough electricity to power 1,000 homes for a year, and emitted hundreds of tons of CO₂." Flag this precisely: OpenAI has never disclosed GPT-4's training energy use, so this is an unverified, widely-recirculated internet estimate, not a confirmed figure — treat it as illustrative of scale, not as data. It's also worth separating this from the hook's "five cars" figure: that number comes from a 2019 academic study of a much smaller, older-generation NLP model, not from GPT-4 or any comparably large modern model, so the two shouldn't be read as two data points about the same kind of training run — they're rough estimates for two different models from two different eras. Introduce the framing this age band should own: a full accounting requires lifecycle carbon assessment — training, inference (every subsequent use, which for a widely-deployed model can exceed training's total energy over the model's lifetime), hardware manufacturing (fabricating the specialized chips, called GPUs/accelerators, involves energy-intensive processes and rare materials), the broader supply chain, and end-of-life disposal of retired hardware. Point out that headline figures almost always report only the first item, because it's the easiest to measure and the most dramatic-sounding, which is itself worth flagging as a critical-reading skill. On "Data Centers" (1-2% of global electricity, doubling every few years — a recent-estimates snapshot that gets revised often as AI adoption changes, not a fixed number): introduce carbon intensity — the same kilowatt-hour of electricity can represent wildly different CO₂ emissions depending on the local grid mix (a data center in a hydro- or nuclear-heavy grid emits far less per unit of compute than one on a coal-heavy grid), which is why "electricity used" and "CO₂ emitted" are related but not interchangeable numbers. On "Water Usage" (Google's AI: 5.6 billion gallons in 2023): correct the framing before going further — that figure is Google's total disclosed data-center water use for 2023, covering search, Gmail, cloud services, and every other workload alongside AI, not an AI-specific number; Google has not published a breakout isolating AI's share. Use it to distinguish water withdrawal (taken from a source) from water consumption (not returned, largely lost to evaporation in cooling) — a distinction real corporate sustainability reports increasingly have to make, and one a skeptical reader should ask about whenever a water figure is quoted without specifying which one it is. Generalize the attribution problem from there: when a company does disclose a genuinely AI-specific water or energy number, it's usually derived from an internal apportionment of shared infrastructure costs between AI and non-AI workloads, and that methodology is rarely published in full detail — so even an "AI-labeled" figure deserves the same skepticism as this mislabeled one.
Scene 2 — AI Helping the Planet. "AI optimizes power grids, reducing energy waste by up to 15% — saving more energy than AI consumes." Have students identify the comparison being made and its limits: this is a claim about grid-optimization applications specifically, not a company-wide or industry-wide accounting, and "saves more than AI consumes" requires defining the boundary of what counts as "AI" in each half of that comparison — a classic scope problem in any cost-benefit claim. On "Smart Farming" (90% pesticide reduction via AI-monitored drones): connect to carbon-aware computing, a real, deployed practice where compute-intensive but flexible workloads (including some AI training and batch inference jobs) are scheduled to run when and where the electricity grid's carbon intensity is lowest — e.g., shifting a training run to hours when a region's grid has more solar or wind online — rather than running full-tilt regardless of the grid's instantaneous carbon cost. On "Ocean Cleanup": use this to discuss the limits of even a technically sound environmental-AI application — detection is necessary but not sufficient; the actual cleanup still requires funded ships, crews, and disposal infrastructure, so an AI system's real-world impact depends on the non-AI systems around it. This is a useful general pattern to name: an AI system's headline capability (detecting plastic, monitoring crop stress, forecasting energy demand) is often the easy part relative to the institutional and financial infrastructure required to act on what it detects — worth remembering before crediting AI itself with an environmental outcome that actually depended on a much larger surrounding system.
Scene 3 — The Balance. Spend the most time here; it's the intellectual core of the lesson for this age band. "Net Impact: does AI's environmental benefit outweigh its energy cost? Researchers debate this actively." On "Green AI" (models using 100x less energy for similar results): introduce Mixture-of-Experts (MoE) architectures precisely — a model is built from many specialized sub-networks ("experts"), and a routing mechanism sends each individual input to only a small subset of them, so the model's total parameter count (and capability) can be enormous while the computation actually performed per request stays much smaller than running a comparably capable dense model. This is a real, deployed architecture family, not a simplification for the lesson. On "Solar Power": note that renewable-powered data centers address the carbon-intensity side of the equation but not the absolute electricity demand or water-for-cooling side — a fully solar-powered data center can still have a large water footprint. On "Carbon Labels": connect to the rebound effect (Jevons paradox) as the lesson's final and most important idea — historically, when a resource-consuming technology becomes more efficient, the resulting drop in cost per use tends to drive increased total usage, sometimes enough to offset or exceed the efficiency gain. Applied to AI: a 100x more efficient model may get deployed 1000x more widely once it's cheap enough, and the honest, current answer is that nobody has definitively shown which effect wins at an industry-wide scale. This is a good point to note the pattern's history: the rebound effect isn't an AI-specific idea — it was first described in the 1800s for coal-fired steam engines becoming more fuel-efficient, and versions of it have shown up in lighting (LED bulbs made light so cheap that total lighting energy use didn't fall as much as the efficiency gain alone would predict) and computing generally. AI is simply the newest, fastest-moving case of an old pattern.
A concrete case for weighing this tradeoff: the Philippines. Take everything covered above and apply it to one real country rather than leaving it abstract. The Philippines is consistently ranked among the world's most climate-vulnerable nations — it sits inside the most active tropical cyclone basin on Earth, absorbing roughly 20 named storms a year through the Philippine Area of Responsibility, several of them destructive enough to cause mass casualties and billions of pesos in damage (2013's Typhoon Yolanda/Haiyan is the reference point most students will already know). PAGASA, the national weather service, and partner efforts — including AI-based hydrological forecasting tools such as Google's Flood Hub, which now models river basins across the Philippines — increasingly rely on machine-learning models trained on historical weather, satellite, and river-gauge data to extend warning lead times for typhoons and flash floods. That is a genuine instance of this lesson's "AI helping the planet" side, not a hypothetical: more accurate, earlier warnings translate directly into evacuation time, which is the single biggest lever for reducing a storm's death toll.
Now put that next to the cost side, honestly. Any AI infrastructure serving the Philippines — a local data center, or a cloud region abroad handling Philippine traffic — carries the same electricity and water costs covered in Scene 1, and the Philippines already experiences seasonal water stress in parts of the country, so Scene 1's water-footprint question isn't hypothetical here either. Push students toward the real nuance: this isn't a case where the cost and the benefit simply cancel out, because they aren't the same kind of AI system running at the same scale in the same place. A national weather service's forecasting models represent a tiny fraction of global AI compute; a large consumer chatbot's training run does not. Treating "AI's environmental impact" as one undifferentiated number — which this lesson's own headline figures can tempt you to do — obscures the fact that a climate-vulnerable country can rationally want more of one kind of AI (small, targeted, disaster-forecasting models) while remaining a legitimate critic of the environmental cost of another kind (massive general-purpose models, comparatively little of whose compute goes toward local climate resilience). Ask directly: "Does it make sense to evaluate 'AI's environmental impact' as a single number, or does this example show why it has to be broken down by application?"
Close by returning to the opening question: "Earlier I asked where you'd start if you were an engineer told to reduce your product's environmental footprint. Given everything we've covered — lifecycle assessment, carbon intensity, carbon-aware scheduling, MoE architectures, and the rebound effect — revise or defend your original answer." This closes the loop between the hook and the lesson content and makes the technical vocabulary functional rather than decorative.
Extension activity (25-30 minutes, can extend to homework): Assign a short position paper (roughly 300-500 words): "Should governments require AI companies to publish standardized environmental-impact disclosures for their models, similar to vehicle fuel-economy labels?" Students must address at least: what should be measured (training only, or full lifecycle), who would verify the disclosures, one argument for mandatory disclosure, and one genuine counterargument (e.g., competitive/trade-secret concerns about revealing model architecture details, or the practical difficulty of standardizing a measurement across very different model types). Grade for whether the argument engages with the actual technical tradeoffs covered in class, not just general environmentalist or techno-optimist sentiment. As a stretch option for students considering computing-related fields, have them research one real company's published sustainability or environmental report for its AI or cloud division and evaluate, using the lifecycle-assessment framework from class, what the report does and doesn't disclose. Students more interested in the resilience side than the cost side may substitute a real AI-based disaster-forecasting or climate-adaptation program — such as an AI-assisted flood or typhoon forecasting tool used in the Philippines — and evaluate what it discloses about its data sources, accuracy, and limitations, using the same critical-evaluation habits from Scene 1's fact-checking.
You can use, adapt and share this guide for non-commercial purposes under CC BY-NC-SA 4.0, with credit to AIQ. Copy this line to give credit: