Quiz: Environmental Data Ethics and Climate

Test your understanding before moving to the next chapter. Target: 70% or higher to proceed.

Section 1: Multiple Choice (1 point each)

1. Approximately what percentage of global electricity consumption did data centers account for in 2024?

  • A) 0.2%
  • B) 2%
  • C) 10%
  • D) 20%
Answer **B)** 2%. *Explanation:* Section 34.1.1 states that data centers consumed approximately 460 TWh globally in 2024, representing approximately 2% of global electricity consumption — roughly equivalent to the total electricity consumption of France. This figure is growing rapidly as AI workloads increase.

2. Carbon intensity varies by region because:

  • A) Different regions use different programming languages.
  • B) The electricity grid in each region relies on different mixes of energy sources, from low-carbon (hydro, nuclear, wind) to high-carbon (coal, natural gas).
  • C) Some regions have more efficient algorithms.
  • D) Carbon emissions are only produced by servers, not by cooling systems.
Answer **B)** The electricity grid in each region relies on different mixes of energy sources, from low-carbon (hydro, nuclear, wind) to high-carbon (coal, natural gas). *Explanation:* Section 34.1.2 explains that carbon intensity measures grams of CO2 per kilowatt-hour and varies dramatically by region — from ~1.2 gCO2/kWh in Quebec (99%+ hydroelectric) to ~900 gCO2/kWh in South Africa (coal). The same computation can produce vastly different carbon emissions depending on where it is performed, making location a governance choice.

3. The CarbonEstimator class calculates total energy by accounting for:

  • A) GPU power draw only.
  • B) GPU power draw adjusted by utilization, multiplied by number of GPUs and training hours, then multiplied by PUE for data center overhead.
  • C) The total cost of electricity in dollars.
  • D) The weight of the hardware in kilograms.
Answer **B)** GPU power draw adjusted by utilization, multiplied by number of GPUs and training hours, then multiplied by PUE for data center overhead. *Explanation:* Section 34.2.3 traces through the calculation: (1) GPU power in kW, (2) adjusted by utilization factor (e.g., 0.7), (3) multiplied by number of GPUs, (4) multiplied by training hours to get GPU energy in kWh, (5) multiplied by PUE to account for cooling and other data center overhead. This produces total facility energy, which is then multiplied by regional carbon intensity to get CO2 emissions.

4. Training a large language model (2,048 H100 GPUs for 30 days) in Quebec rather than Virginia would reduce carbon emissions by approximately:

  • A) 10%
  • B) 50%
  • C) 75%
  • D) 91%
Answer **D)** 91%. *Explanation:* Section 34.2.3 calculates that training the same model in Quebec (carbon intensity ~30 gCO2/kWh) rather than Virginia (340 gCO2/kWh) reduces carbon emissions by approximately 91%. This dramatic difference demonstrates why data center location is one of the most significant governance levers for reducing AI's carbon footprint.

5. The Green AI movement advocates for:

  • A) Banning all AI research until renewable energy sources power 100% of the grid.
  • B) Efficiency as a first-class metric, reporting compute budgets, and prioritizing research accessible to resource-constrained institutions.
  • C) Training only the largest possible models to maximize accuracy.
  • D) Relocating all data centers to Antarctica for natural cooling.
Answer **B)** Efficiency as a first-class metric, reporting compute budgets, and prioritizing research accessible to resource-constrained institutions. *Explanation:* Section 34.4.1 describes Green AI (Schwartz et al., 2020) as advocating for three practices: (1) treating efficiency (accuracy per computation) as a first-class metric alongside raw accuracy, (2) publishing computational resources used in model development, and (3) prioritizing research that produces useful results within budgets accessible to academic researchers, small companies, and researchers in the Global South. Green AI stands in contrast to "Red AI" — the trend toward ever-larger models at enormous computational cost.

6. The "rebound effect" (Jevons paradox) in the context of AI efficiency means:

  • A) More efficient models always reduce total energy consumption.
  • B) Efficiency gains reduce per-unit computation costs, potentially leading to increased demand that offsets or exceeds the savings.
  • C) Rebounding from errors makes models more accurate.
  • D) Data centers become less efficient over time.
Answer **B)** Efficiency gains reduce per-unit computation costs, potentially leading to increased demand that offsets or exceeds the savings. *Explanation:* Section 34.4.3 warns that if more efficient models make AI cheaper, organizations may train more models, run more experiments, and deploy AI in more contexts — potentially increasing total emissions even as per-model emissions decline. This means technical efficiency alone cannot solve the environmental challenge; governance mechanisms (carbon reporting, emissions caps, carbon pricing) may be necessary to ensure efficiency gains translate into actual reductions.

7. Li et al. (2023) estimated that training GPT-3 consumed approximately how much fresh water?

  • A) 7,000 liters
  • B) 70,000 liters
  • C) 700,000 liters
  • D) 7,000,000 liters
Answer **C)** 700,000 liters. *Explanation:* Section 34.1.3 cites the study estimating that training GPT-3 consumed approximately 700,000 liters of fresh water for cooling — enough to fill about 370 standard bathtubs. Water consumption is a critical environmental justice issue because data centers are often located in regions experiencing water stress.

8. Knowledge distillation, as a Green AI technique, involves:

  • A) Removing all parameters from a large model.
  • B) Training a smaller "student" model to mimic a larger "teacher" model, achieving comparable performance with fewer parameters.
  • C) Distilling water for data center cooling.
  • D) Converting unstructured data into structured data.
Answer **B)** Training a smaller "student" model to mimic a larger "teacher" model, achieving comparable performance with fewer parameters. *Explanation:* Section 34.4.2 describes knowledge distillation as one of several model compression techniques that can reduce model size by 50-90% with minimal accuracy loss. By transferring the "knowledge" of a large model to a smaller one, distillation achieves most of the performance benefit at a fraction of the computational and environmental cost.

9. The chapter identifies the "dual role" of data systems in relation to the environment. This means:

  • A) Data systems are both expensive and cheap to operate.
  • B) Data systems contribute to environmental harm through energy consumption and e-waste while also enabling essential environmental monitoring, climate modeling, and conservation.
  • C) Data systems can run on both renewable and non-renewable energy.
  • D) Data systems produce both structured and unstructured data.
Answer **B)** Data systems contribute to environmental harm through energy consumption and e-waste while also enabling essential environmental monitoring, climate modeling, and conservation. *Explanation:* Section 34.5 presents this dual role: data systems are environmental costs (energy, carbon, water, e-waste) and environmental tools (satellite observation, sensor networks, climate modeling, conservation AI). This duality complicates simple narratives — you cannot argue for eliminating data infrastructure without acknowledging the environmental monitoring it enables, nor can you ignore its costs by pointing to its benefits.

10. Environmental data justice, as described in Section 34.6, requires:

  • A) Building more data centers in every country.
  • B) Transparency about environmental costs, accountability through carbon pricing, participation by affected communities, and equitable distribution of both benefits and costs.
  • C) Eliminating all AI research until climate change is solved.
  • D) Locating all data centers in the Global South because land is cheaper.
Answer **B)** Transparency about environmental costs, accountability through carbon pricing, participation by affected communities, and equitable distribution of both benefits and costs. *Explanation:* Section 34.6.2 identifies four requirements for environmental data justice: transparency (about carbon, water, e-waste, and supply chain impacts), accountability (through carbon pricing and extended producer responsibility), participation (by affected communities in siting and operations decisions), and equitable distribution (of both benefits and costs of data systems).

Section 2: True/False with Justification (1 point each)

11. "The CarbonEstimator class provides a precise and comprehensive accounting of all carbon emissions associated with model training."

Answer **False.** *Explanation:* Section 34.2.2's callout box explicitly states that the estimator provides a "reasonable lower bound" and does not account for embodied carbon in GPU manufacturing, network/storage energy beyond PUE, failed training runs, hyperparameter search, inference emissions, or water consumption. The estimator is a starting point for making environmental costs visible, not a comprehensive carbon accounting tool.

12. "OpenAI publicly disclosed the carbon footprint of training GPT-4."

Answer **False.** *Explanation:* Section 34.3.3 explicitly states that "OpenAI did not disclose the carbon footprint of training GPT-4." Independent estimates range from 3,000 to 13,000 tonnes of CO2. The lack of disclosure is itself identified as a governance issue — if environmental impact is not measured and reported, it cannot be governed.

13. "Carbon-aware scheduling — timing training runs to coincide with periods of high renewable energy generation — can reduce emissions without changing total computation."

Answer **True.** *Explanation:* Section 34.4.2 identifies carbon-aware scheduling as a Green AI technique. By scheduling training runs during periods when the electricity grid has lower carbon intensity (e.g., midday when solar generation peaks, windy periods for wind-heavy grids), the same computation produces fewer emissions. The total computation remains identical; the timing changes the carbon intensity of the electricity consumed.

14. "The environmental costs of data infrastructure are distributed equitably across communities worldwide."

Answer **False.** *Explanation:* Section 34.6.1 documents four patterns of inequitable distribution: data center siting in communities with less political power, e-waste exported to the Global South, rare earth mining concentrated in developing countries under exploitative conditions, and climate impacts falling disproportionately on the Global South. The benefits of AI accrue to wealthy companies and their users; the environmental costs are externalized to communities with the least power to resist.

15. "Transfer learning and fine-tuning can reduce the environmental cost of AI by distributing the carbon footprint of pre-training across many downstream applications."

Answer **True.** *Explanation:* Section 34.4.2 describes transfer learning and fine-tuning as approaches that achieve strong performance with a fraction of the training cost of training from scratch. By building on pre-trained models, downstream applications share the environmental cost of the original pre-training — making the per-application environmental footprint significantly smaller than if each application required training from scratch.

Section 3: Short Answer (2 points each)

16. Explain the concept of "stranded assets" as it could apply to data center infrastructure in the context of climate change. What happens if carbon pricing or renewable energy requirements make existing fossil-fuel-powered data centers uneconomical?

Sample Answer Stranded assets are investments that lose their value before the end of their expected economic life due to changes in regulations, market conditions, or physical circumstances. In the data center context, if governments impose carbon pricing, mandatory renewable energy requirements, or emissions caps, data centers powered by fossil fuels in high-carbon regions could become stranded assets — their operating costs would increase dramatically, their competitiveness would decline relative to data centers in low-carbon regions, and the capital invested in their construction would be partially or fully lost. This creates a governance tension: data center operators may resist climate regulations because of stranded asset risk, while the environmental costs of continued fossil-fuel-powered operation accumulate. Proactive investment in renewable-powered data centers and energy efficiency reduces stranded asset risk while addressing environmental concerns — but requires governance signals (carbon pricing, clean energy standards) to shift investment incentives before assets become stranded. *Key points for full credit:* - Defines stranded assets correctly - Applies the concept to data center infrastructure specifically - Identifies the governance implications (resistance to regulation vs. proactive investment)

17. The chapter argues that Strubell et al.'s (2019) most lasting contribution was not the specific carbon numbers but the argument that carbon emissions should be reported alongside accuracy metrics. Explain why this reporting norm matters and what it would change about AI research culture.

Sample Answer Requiring carbon emissions reporting alongside accuracy metrics would transform AI research culture by making environmental cost visible and subject to evaluation. Currently, AI papers report model accuracy, precision, recall, and other performance metrics — creating incentives to maximize these metrics regardless of computational cost. If carbon emissions were also reported, researchers and reviewers would be able to evaluate whether marginal accuracy gains justify their environmental cost. A model that achieves 96% accuracy at 10x the carbon footprint of a 95%-accurate model would face scrutiny that currently does not exist. This reporting norm would also incentivize research into efficient methods (Green AI), enable fair comparison between approaches, support the development of carbon budgets for research projects, and create accountability for the environmental consequences of research choices. Most fundamentally, it would shift the culture from one that treats computation as free (someone else pays the electricity bill) to one that recognizes computation as having real, measurable environmental costs that are part of the evaluation of research quality. *Key points for full credit:* - Explains the current incentive structure (accuracy maximization without cost consideration) - Describes how reporting would change incentives and culture - Connects to Green AI principles (efficiency as a first-class metric)

18. Explain the environmental justice dimension of e-waste from AI hardware. Who generates the e-waste, and who bears the health and environmental consequences?

Sample Answer AI hardware (GPUs, TPUs) is manufactured by companies in the US, Taiwan, and South Korea, used by technology companies primarily in the US, Europe, and East Asia, and retired after 3-5 year lifespans as newer, more efficient hardware becomes available. The resulting e-waste contains toxic materials — lead, mercury, cadmium, brominated flame retardants. Only approximately 22% of global e-waste is formally collected and recycled; the remainder is landfilled, incinerated, or exported to countries in the Global South — primarily Ghana, Nigeria, India, and China. In these countries, informal recycling operations expose workers (including children) to toxic substances as they extract valuable metals from circuit boards using acid baths, open burning, and other hazardous methods. The environmental justice dimension is stark: the communities that generate the e-waste (wealthy technology companies and their customers) bear none of its health consequences, while the communities that process it (low-income workers in the Global South) bear the full burden of toxic exposure, soil contamination, and water pollution. The benefits flow north; the waste flows south. *Key points for full credit:* - Traces the lifecycle from manufacture through use to disposal - Identifies the geographic distribution (generation in Global North, disposal in Global South) - Explains the health consequences for communities processing e-waste - Frames the disparity as an environmental justice issue

19. Using the CarbonEstimator class, explain the step-by-step calculation for estimating the carbon footprint of training a model on 16 A100 GPUs for 120 hours in eu-north. What is the approximate result, and why is it dramatically lower than the same calculation in us-east?

Sample Answer Step-by-step calculation: 1. **GPU power:** A100 TDP = 400W = 0.4 kW 2. **Utilization adjustment:** 0.4 kW x 0.7 (default utilization) = 0.28 kW per GPU 3. **Total GPU energy:** 0.28 kW x 16 GPUs x 120 hours = 537.6 kWh 4. **PUE adjustment:** 537.6 kWh x 1.1 (default PUE) = 591.36 kWh 5. **Carbon intensity (eu-north):** 25 gCO2/kWh = 0.025 kgCO2/kWh 6. **Total carbon:** 591.36 x 0.025 = 14.78 kg CO2 The same calculation in us-east (340 gCO2/kWh = 0.34 kgCO2/kWh): 591.36 x 0.34 = 201.06 kg CO2 The eu-north result (~15 kg CO2) is approximately 13.6 times lower than us-east (~201 kg CO2) because eu-north (Sweden/Finland) has carbon intensity of only 25 gCO2/kWh (powered by hydro, nuclear, and wind) versus us-east's 340 gCO2/kWh (powered by natural gas and nuclear). The energy consumed is identical; the difference is entirely in the carbon intensity of the electricity grid. This demonstrates that location choice is one of the most powerful levers for reducing AI's carbon footprint. *Key points for full credit:* - Traces through all calculation steps correctly - Computes both results - Explains the difference in terms of regional carbon intensity - Connects to the governance implication (location as a choice)

Section 4: Applied Scenario (5 points)

20. Read the following scenario and answer all parts.

Scenario: GreenModel Inc.

GreenModel Inc. is an AI startup developing a climate prediction model for agricultural planning. The model helps farmers in Sub-Saharan Africa predict drought conditions three months in advance, enabling better crop selection and water management. Independent assessments suggest the model has saved an estimated $50 million in crop losses over two years.

To develop the model, GreenModel trains on 512 H100 GPUs for 21 days in a data center in Virginia (us-east, carbon intensity 340 gCO2/kWh). The training data includes satellite imagery, weather station data, and digitized indigenous agricultural knowledge collected from farming communities in Kenya and Ghana — knowledge that was shared in exchange for free access to the final prediction tool.

GreenModel has been nominated for a "Green Tech" award. A journalist investigating the nomination asks three questions: (1) What is the carbon footprint of training the model? (2) Who provided the indigenous agricultural knowledge, and what governance mechanisms protect their interests? (3) Where does the hardware end up when it is retired?

(a) Using the CarbonEstimator class (or manual calculation with the parameters from Section 34.2.2), estimate the carbon footprint of GreenModel's training run. Express it in kg CO2, tonnes CO2, and transatlantic flight equivalents. (1 point)

(b) Recalculate assuming GreenModel had trained in canada-central instead of us-east. How much carbon would have been saved? Express both as absolute reduction and percentage reduction. (1 point)

(c) Apply the CARE Principles (Chapter 32) to the indigenous agricultural knowledge component. Evaluate GreenModel's arrangement (knowledge shared in exchange for free tool access) against each principle. Is the arrangement adequate? What additional governance mechanisms would you recommend? (1 point)

(d) Apply the Environmental Impact Assessment framework from Section 34.7 (Measure, Compare, Contextualize, Mitigate, Distribute, Report) to GreenModel's situation. Walk through each step. (1 point)

(e) Should GreenModel receive the "Green Tech" award? Write a 100-150 word evaluation that weighs the agricultural benefits against the environmental costs, the indigenous knowledge governance concerns, and the environmental justice dimensions. Reference specific concepts from Chapters 32 and 34. (1 point)

Sample Answer **(a)** Carbon footprint calculation: - GPU power: H100 = 700W = 0.7 kW - Utilization: 0.7 x 0.7 = 0.49 kW per GPU - Total GPU energy: 0.49 x 512 x (21 x 24 = 504 hours) = 126,451.2 kWh - PUE: 126,451.2 x 1.1 = 139,096.3 kWh - Carbon (us-east, 0.34 kgCO2/kWh): 139,096.3 x 0.34 = 47,292.7 kg CO2 - Approximately **47.3 tonnes CO2** - Transatlantic flights: 47,293 / 1,600 = approximately **29.6 flights** **(b)** Recalculated for canada-central (0.03 kgCO2/kWh): 139,096.3 x 0.03 = 4,172.9 kg CO2 = approximately **4.2 tonnes CO2** Absolute reduction: 47,293 - 4,173 = 43,120 kg CO2 saved Percentage reduction: 43,120 / 47,293 = approximately **91.2%** reduction **(c)** CARE analysis: - **Collective Benefit:** Partially met. Communities receive free tool access, which provides benefit. But the model also generates revenue for GreenModel (from other customers), and the indigenous knowledge contributes to that commercial value without revenue sharing. - **Authority to Control:** Not met. The knowledge was "shared" (implying transfer of control). No governance mechanism ensures indigenous communities can control how their knowledge is used in the model or in future applications. - **Responsibility:** Partially met. GreenModel provides the tool, but there is no reporting on how the indigenous knowledge is used, no ongoing engagement, and no mechanism for communities to evaluate the arrangement. - **Ethics:** Partially met. The free tool provides genuine benefit, but the exchange of centuries-old ecological knowledge for access to a single tool may not represent equitable value exchange. Additional governance mechanisms needed: community advisory board with decision-making power, benefit-sharing agreement for commercial applications, data sovereignty clause ensuring communities retain governance rights over their knowledge, and regular reporting to contributing communities. **(d)** Environmental Impact Assessment: 1. **Measure:** 47.3 tonnes CO2, ~139 MWh energy, training in us-east. 2. **Compare:** Training in canada-central would reduce emissions by 91%. Using H100 GPUs instead of older hardware reduces training time (and energy) vs. equivalent computation on V100s. 3. **Contextualize:** 47.3 tonnes = ~30 transatlantic flights, ~11 years of average US household electricity, ~2,150 trees needed for one year of offset. 4. **Mitigate:** Relocate training to low-carbon region. Apply model compression for inference efficiency. Schedule training during low-carbon periods. 5. **Distribute:** Environmental costs borne by Virginia community (power plant emissions, water consumption). Benefits accrue to Sub-Saharan African farmers and GreenModel shareholders. The distribution is unequal. 6. **Report:** GreenModel should publicly disclose training emissions, compare to mitigation alternatives, and include environmental impact alongside model performance in publications. **(e)** GreenModel's agricultural benefits are significant and real — $50 million in saved crop losses directly serves vulnerable farming communities. But a "Green Tech" award should require Green practices, not just green outcomes. GreenModel trained in a high-carbon region when a 91% reduction was available by relocating to Canada. Its indigenous knowledge arrangement lacks the governance mechanisms the CARE Principles require — creating a data colonialism risk where African communities' knowledge generates commercial value without adequate sovereignty protections. The award should be conditional on: (1) committing to low-carbon training for future iterations, (2) establishing CARE-compliant governance for indigenous knowledge, and (3) publishing transparent environmental impact reporting. Green outcomes do not excuse extractive practices.

Scoring & Review Recommendations

Score Range Assessment Next Steps
Below 50% (< 15 pts) Needs review Re-read Sections 34.1-34.2 carefully, run CarbonEstimator examples
50-69% (15-20 pts) Partial understanding Review Part B exercises, run Python calculations
70-85% (21-25 pts) Solid understanding Ready to proceed to Chapter 35
Above 85% (> 25 pts) Strong mastery Proceed to Chapter 35: Children, Teens, and Digital Vulnerability
Section Points Available
Section 1: Multiple Choice 10 points (10 questions x 1 pt)
Section 2: True/False with Justification 5 points (5 questions x 1 pt)
Section 3: Short Answer 8 points (4 questions x 2 pts)
Section 4: Applied Scenario 5 points (5 parts x 1 pt)
Total 28 points