Chapter 31: Exercises — The Environmental Cost of AI

Individual Reflection Exercises

Exercise 1: Your AI Carbon Footprint Estimate your personal AI usage for one week: number of ChatGPT, Copilot, Claude, or other AI queries; AI-assisted searches; AI image generations; AI voice assistant queries. Using the approximate figure of 0.003–0.01 kWh per AI query and a grid carbon intensity of 200g CO2/kWh, calculate your weekly AI carbon footprint. Compare to other activities in your carbon footprint (driving a car, flying, eating beef). Write a 400-word reflection on proportionality and individual vs. collective responsibility.

Exercise 2: The Disclosure Audit Visit the sustainability or environmental responsibility pages of OpenAI, Anthropic, Google DeepMind, and Meta AI. For each, document: what environmental metrics are disclosed, at what level of granularity, with what verification, and against what targets. Write a comparative assessment of disclosure quality. Grade each company A–F on environmental transparency and justify your grades.

Exercise 3: Water Stress Mapping Using the World Resources Institute's Aqueduct Water Risk Atlas (wri.org/aqueduct), identify the 10 largest data center clusters in the United States. For each, note the water stress level of the region (low, medium-high, high, extremely high). What percentage of major data center capacity is located in water-stressed areas? Write a 400-word analysis of the geographic risk.

Exercise 4: The Efficiency Trade-off Research a specific pair of AI models that perform similar tasks: a large frontier model (e.g., GPT-4) and a smaller efficient alternative (e.g., a fine-tuned Llama model, or DistilBERT vs. BERT). Find or estimate the difference in computational cost (parameters, FLOP requirements, or inference latency). Research the performance difference on a relevant benchmark. Write a 500-word analysis of whether the performance gain justifies the computational cost.

Exercise 5: The Rebound Effect Analysis Identify a domain in which AI has made a specific task significantly cheaper or faster (e.g., code generation, content writing, data analysis, drug discovery). Research whether total activity in that domain has increased as a result. Write a 500-word analysis of whether the Jevons paradox is operating in this domain — whether efficiency gains have led to increased total resource consumption.

Group Discussion Exercises

Exercise 6: The Data Center Siting Decision Your company is planning to build a new data center to support AI workloads. You have three candidate sites: (a) Phoenix, Arizona — cheap power, high water stress, excellent connectivity; (b) Stavanger, Norway — hydropower with near-zero carbon, abundant water, higher land costs; (c) Singapore — excellent connectivity, near the Southeast Asian market, moderate energy cost, moderate water stress. Make the siting decision as a group, weighting financial, environmental, and ethical considerations. Present and defend your choice.

Exercise 7: The Sustainability Report Review Obtain Google's or Microsoft's most recent environmental sustainability report. Divide into groups: one analyzing the energy and carbon disclosure, one analyzing water disclosure, one analyzing hardware supply chain disclosure. Each group prepares a critical assessment of completeness, accuracy, and comparability. As a class, synthesize: what is well-disclosed, what is missing, and what regulatory requirements would improve it?

Exercise 8: The Green AI Research Agenda You are the research director of an AI lab that has decided to prioritize Green AI practices. Develop a research agenda that: (1) identifies the most promising efficiency improvements for your lab's research area, (2) establishes metrics for reporting computational cost alongside accuracy in publications, (3) creates incentives for researchers to prefer efficient approaches, and (4) engages with the broader AI research community to establish Green AI norms. Present your agenda.

Exercise 9: The Investor Engagement You are an institutional investor holding shares in a major AI company. You want to engage the company's board on its environmental disclosure practices, particularly on Scope 3 emissions from hardware manufacturing and customer use of its models. Develop an engagement strategy: what specific disclosure commitments would you seek, how would you measure progress, and what escalation steps would you take if the company failed to respond?

Exercise 10: The Community Hearing Simulate a public hearing on a proposed 500-acre, 500-megawatt data center in a water-stressed region. Students play: the data center developer, local government officials, agricultural water users, residential community members, an environmental advocacy organization, and a labor union representing construction workers. Hold the hearing, with each group presenting testimony and responding to questions. Vote on whether to approve the conditional use permit.

Analytical Exercises

Exercise 11: Carbon Cost Estimation Using the Patterson et al. (2021) methodology, estimate the training carbon cost of a fictional large language model with the following specifications: 100 billion parameters, trained for 4 weeks, using 5,000 A100 GPUs, at a data center in Virginia with a grid carbon intensity of 300g CO2/kWh. Show your calculations. Then recalculate assuming the data center is powered by 100% hydropower (carbon intensity: 5g CO2/kWh). What does this illustrate about the leverage of renewable energy sourcing?

Exercise 12: WUE Analysis Research the Water Use Effectiveness (WUE) metrics published by Microsoft, Google, and Amazon for their data center fleets. For each company, calculate the total annual water consumption of their data centers using published WUE and total power consumption figures. If any company has not published the required figures, note the gap. Write a comparative analysis.

Exercise 13: Supply Chain Carbon Research the Scope 3 supply chain carbon footprint of NVIDIA's H100 GPU — the chip most widely used for AI training. This requires research into semiconductor manufacturing energy use, rare earth mineral extraction carbon, shipping and logistics, and end-of-life e-waste processing. Document the available information and the gaps. Write a 1,000-word analysis of the supply chain carbon embedded in AI training infrastructure.

Exercise 14: The Regulatory Gap Analysis Compare the environmental disclosure requirements that currently apply to AI companies under: (a) US SEC climate disclosure rules, (b) EU Corporate Sustainability Reporting Directive, (c) EU AI Act environmental provisions, and (d) voluntary frameworks like CDP and the Green Software Foundation SCI. Identify the gaps in each framework and what additional requirements would be needed for comprehensive environmental accountability of AI companies.

Exercise 15: AI Climate Benefit Analysis Select one specific AI application that is claimed to have climate benefits (e.g., DeepMind's AlphaFold for drug discovery, AI-optimized energy grids, AI precision agriculture). Conduct a rigorous analysis: what is the energy and carbon cost of the AI system? What is the measured or credibly estimated climate benefit? What is the counterfactual — what would have happened without the AI system? What is the net climate impact? Document the significant uncertainties in your analysis.

Case Application Exercises

Exercise 16: The OpenAI Environmental Policy Draft a comprehensive environmental policy for OpenAI that would make it a leader in AI environmental accountability. Your policy should address: mandatory environmental disclosure for model training and inference, renewable energy commitments with 24/7 carbon-free energy standard, water use disclosure and efficiency commitments, hardware supply chain standards, e-waste management, and contributions to climate finance proportional to historical emissions. Make your policy specific and implementable.

Exercise 17: The Data Center Water Policy You are advising the state of Arizona on water policy for data centers. Draft a regulatory framework that: addresses the cumulative water impact of data center concentration in water-stressed areas, creates incentives for water-efficient cooling technologies, establishes disclosure requirements for data center water use, provides community input mechanisms for new data center permitting, and balances economic development interests with water resource sustainability. Draft specific regulatory language for at least three provisions.

Exercise 18: The Corporate Carbon Accounting A company (choose a specific company you are familiar with) has deployed Microsoft 365 Copilot for its 10,000 employees, each making an average of 20 AI-assisted queries per day. Calculate the additional Scope 2 and Scope 3 carbon emissions from this AI deployment under: (a) current voluntary reporting standards (likely: not reported), and (b) what a comprehensive Scope 3 disclosure requirement would require. What data would the company need to collect? What would it report? How does this change the company's total reported carbon footprint?

Exercise 19: The Green AI Publication Standard Design a standard for AI research publication that requires reporting computational cost alongside accuracy metrics, analogous to how clinical research requires reporting of statistical methods, sample sizes, and confidence intervals. Your standard should specify: what metrics must be reported, how they should be measured, how they should be normalized for comparison across different hardware, and how journals and conferences should implement and enforce the standard.

Exercise 20: The Efficient AI Deployment Your company is evaluating three options for deploying AI to handle 1 million customer service queries per day: (a) GPT-4 API (high quality, high energy cost); (b) a smaller open-source model fine-tuned on your data (moderate quality, moderate energy cost); (c) a rule-based system for common queries plus AI for complex ones (lower quality for complex queries, lower energy cost). Design the analysis framework for choosing among these options, incorporating performance, cost, and environmental criteria. Make a recommendation and defend it.

Research and Writing Exercises

Exercise 21: The Strubell Effect Research the academic and policy impact of Emma Strubell et al.'s 2019 paper "Energy and Policy Considerations for Deep Learning in NLP." How did the research community respond? Did it change disclosure norms in AI research publications? Did it lead to regulatory or policy responses? Did major AI companies respond to its findings? Write a 1,500-word analysis of what the paper's impact reveals about how scientific research on AI's environmental impact has — and has not — translated into changed practice.

Exercise 22: Global Data Center Water Geography Using publicly available information on major data center locations globally, map the concentration of data center capacity against WRI Aqueduct water stress maps. Identify the regions where the largest concentrations of data center capacity coincide with the highest water stress. Research what, if any, regulatory or corporate responses have addressed these conflicts. Write a 1,200-word geographic analysis of the AI water stress problem.

Exercise 23: The Semiconductor Supply Chain Investigation Research the supply chain for a specific rare earth mineral used in AI chips — choose from cobalt, tantalum, or neodymium. Document: where it is mined, by whom, under what environmental and labor conditions, how it is processed, how it enters the global electronics supply chain, and what the AI industry's end-of-life management for hardware containing this mineral looks like. Write a 1,500-word supply chain investigation.

Exercise 24: Net Zero AI Research what "net zero AI" would actually require — not just carbon offsets and renewable energy certificates, but genuine Scope 1, 2, and 3 net zero for the full AI lifecycle including training, inference, hardware manufacturing, and end-of-life. Identify what technical, operational, and regulatory changes would be necessary. Assess whether it is achievable under current trajectories or what inflection points would be necessary. Write a 1,500-word assessment.

Exercise 25: Community Impact Report Research a specific community adjacent to a major AI data center or data center cluster — choose from Ashburn, Virginia; Phoenix/Mesa, Arizona; Dublin, Ireland; or Groningen, Netherlands. Document: what data centers are present or planned, what their energy and water consumption is (using available public information), what local infrastructure impacts they have, what economic benefits they provide, and what community concerns have been raised. Write a balanced 1,500-word community impact report.