Exercises: AI and the Environment — Climate, Resources, and Sustainability
These exercises progress from concept checks to challenging applications. Estimated completion time: 2–3.5 hours.
Difficulty Guide: - ⭐ Foundational (5–10 min each) - ⭐⭐ Intermediate (10–20 min each) - ⭐⭐⭐ Challenging (20–40 min each) - ⭐⭐⭐⭐ Advanced/Research (40+ min each)
Part A: Conceptual Understanding ⭐
A.1. Explain the difference between the energy consumed during training an AI model and the energy consumed during inference. Which is a one-time cost and which is ongoing? Which is likely to become the larger share of total AI energy consumption as AI usage grows?
A.2. What is Power Usage Effectiveness (PUE)? A data center has a PUE of 1.3. In plain language, what does this mean? If the facility improves its PUE to 1.1, what changes?
A.3. Define "embodied carbon" and explain why it matters for AI hardware. Why would it be misleading to evaluate AI's environmental footprint by looking only at electricity consumption during operation?
A.4. What is e-waste, and why is it relevant to AI's environmental impact? Where does most e-waste end up, and why is this an environmental justice issue?
A.5. Explain the rebound effect (Jevons paradox) in your own words. Then give an everyday example outside of AI where efficiency improvements led to increased total consumption.
A.6. What does "green AI" mean? Name at least three techniques associated with the green AI movement.
Part B: Applied Analysis ⭐⭐
B.1. Estimation exercise: A company trains a model using 1,000 GPUs, each consuming 350 watts, for 14 days (24 hours/day). The data center has a PUE of 1.2, and the regional electricity grid has a carbon intensity of 500 grams of CO2 per kWh.
- Calculate the total energy consumed (in kWh).
- Calculate the estimated CO2 emissions (in kg and metric tons).
- Express the emissions as a car-miles equivalent (use 411 grams CO2 per mile).
- Now recalculate assuming the data center runs on 100% renewable energy (carbon intensity = 0). What is the difference?
B.2. Comparative analysis: Compare the environmental footprint of the following three AI use cases. For each, identify the primary environmental cost (energy, water, hardware, or e-waste) and assess whether the environmental benefit (if any) likely outweighs the cost:
- (a) A generative AI tool that produces marketing copy for an advertising firm
- (b) An AI system that monitors satellite imagery to detect illegal deforestation in real time
- (c) A recommendation algorithm that serves video suggestions to 500 million users
B.3. Source evaluation: A major technology company publishes a sustainability report claiming that its data centers are "100% powered by renewable energy." Based on what you learned in Section 18.6, identify three questions you should ask before accepting this claim at face value.
B.4. Scenario analysis: A university is deciding between two options for deploying an AI-powered tutoring system:
- Option A: Use a cloud-based API powered by a large language model hosted in a data center 2,000 miles away
- Option B: Deploy a smaller, distilled model on a local server at the university
Compare the two options in terms of: energy consumption, carbon emissions, performance quality, and cost. Which would you recommend, and under what conditions might the other option be preferable?
B.5. Life cycle assessment: Sketch a simplified life cycle assessment for a GPU used in AI training. Identify at least five stages in the lifecycle (from mineral extraction through disposal) and describe the primary environmental impact at each stage.
Part C: Research Design & Argumentation ⭐⭐–⭐⭐⭐
C.1. Policy argument: Write a 400-word argument either for or against the following proposal: "Companies that train AI models above a certain size threshold should be required to purchase carbon offsets equal to the estimated emissions of the training process." Address at least two counterarguments to your position.
C.2. Framework application: Apply the Evidence Evaluation framework from Section 18.1 to the following claim: "Training GPT-4 produced as much carbon emissions as 1,000 transatlantic flights." Research the claim (or use the information in this chapter) and evaluate: Is it based on training or inference? What energy source is assumed? Is the comparison meaningful or misleading?
C.3. Technical analysis: Research one "green AI" technique (knowledge distillation, pruning, quantization, or efficient architecture design) in more depth. Write a 300-word explanation suitable for a non-technical reader that covers: what the technique does, how much it can reduce energy consumption, what trade-offs are involved, and where it is already being used in practice.
C.4. Design challenge: Design a "carbon label" for AI products — similar to energy efficiency labels on appliances or nutritional labels on food. What information should the label include? How should it be calculated? Who should verify it? Create a mock-up of what the label might look like.
Part D: Synthesis & Critical Thinking ⭐⭐⭐
D.1. Cross-chapter integration: In Chapter 10, you examined the economic forces driving AI adoption. In this chapter, you learned about AI's environmental costs. Analyze the tension between these two forces: economic incentives push for rapid AI scaling, while environmental sustainability may require restraint. Propose a specific policy mechanism that could align economic incentives with environmental goals without stifling beneficial AI development.
D.2. Critique: A technology executive argues: "We should not slow down AI development because of environmental concerns. AI will help us solve climate change, and the benefits far outweigh the costs." Evaluate this argument. What is its strongest point? What does it assume? What evidence would you need to test the claim that "benefits outweigh costs"?
D.3. Transfer: The rebound effect is not unique to AI — it has been observed in transportation (more fuel-efficient cars leading to more driving), lighting (efficient LED bulbs leading to more illumination), and other domains. Choose one non-AI example of the rebound effect and compare it to the AI case. What structural factors determine whether the rebound effect is partial (efficiency gains still produce some net environmental benefit) or total (all gains are consumed by increased use)?
D.4. The paradox question: This chapter presented AI as both an environmental problem and an environmental solution. Write a 500-word essay addressing the question: "Is it possible for AI to have a net positive environmental impact? Under what conditions?" Your essay should consider both the direct environmental costs of AI and its potential contributions to climate mitigation, and should address the rebound effect.
Part M: Mixed Practice (Interleaved) ⭐⭐–⭐⭐⭐
These problems require you to draw on concepts from multiple chapters.
M.1. (From Chapter 3) In Chapter 3, you learned about the training process — how models adjust their parameters through iterations over data. How does the number of parameters in a model relate to the energy required for training? Why have researchers been building larger and larger models despite the environmental cost?
M.2. (From Chapter 5) Large language models like GPT-4 and Claude are trained on enormous datasets and have hundreds of billions of parameters. Using what you know from Chapter 5 about how these models work and from this chapter about their environmental cost, explain why a single LLM query uses significantly more energy than a traditional web search.
M.3. (From Chapter 9) Consider the environmental justice dimension of AI. Using concepts from Chapter 9 (bias, fairness, and the question of "who benefits, who is harmed") and this chapter, analyze whether AI's environmental footprint constitutes a form of environmental injustice. Who benefits from AI systems, and who bears the environmental costs?
M.4. (From Chapter 13) Using the governance frameworks from Chapter 13, design a regulatory approach for managing AI's environmental impact. Should this be handled through existing environmental regulation, new AI-specific regulation, industry self-regulation, or some combination? Justify your approach.
Part E: Research & Extension ⭐⭐⭐⭐
E.1. Data investigation: Research the environmental commitments of three major AI companies (e.g., Google, Microsoft, Meta, Amazon, OpenAI). For each, document: (a) their stated environmental goals, (b) their reported progress, (c) whether their emissions have increased or decreased in recent years, and (d) any discrepancies between their claims and their actual performance. Write a 500-word comparative analysis.
E.2. Environmental impact of your AI use: For one week, track your personal use of AI-powered services (search engines, chatbots, recommendation algorithms, voice assistants, generative AI tools). Estimate the number of queries/interactions per day. Using the rough estimates from this chapter (e.g., a ChatGPT query uses roughly 10x the energy of a Google search), calculate a rough estimate of the carbon footprint of your personal AI usage. Reflect: does the result surprise you? Would it change your behavior?
E.3. Deep dive: Read the original "Green AI" paper by Schwartz et al. (2019, arXiv:1907.10597) or the Strubell et al. (2019) paper on energy and policy considerations for deep learning. Write a 750-word response evaluating the authors' arguments in light of developments since publication. Have their concerns been addressed? Have they become more or less urgent?
Solutions
Selected solutions in appendices/answers-to-selected.md.