Quiz: AI and the Environment — Climate, Resources, and Sustainability

Test your understanding before moving on. Target: 70% or higher to proceed confidently.

Section 1: Multiple Choice (1 point each)

1. The landmark 2019 study by University of Massachusetts Amherst researchers found that training a single large NLP model could produce carbon emissions roughly equivalent to:

  • A) One transatlantic flight
  • B) Five cars over their entire lifetimes (including manufacturing)
  • C) Heating a house for one month
  • D) Running a data center for one year
Answer **B)** Five cars over their entire lifetimes (including manufacturing) *Why B:* The study by Strubell et al. estimated approximately 284 metric tons of CO2 for training a large Transformer with neural architecture search, comparable to five average American cars over their full lifespans. *Why not A:* A single transatlantic flight produces roughly 1–3 tons of CO2 per passenger — far less than the study's estimate. *Why not C:* Heating a house for one month produces a few hundred kg of CO2, orders of magnitude less. *Why not D:* A data center consumes far more energy annually than a single model training run. *Reference:* Section 18.1

2. Which typically consumes more total energy for a widely used AI system over its lifetime?

  • A) Training the model
  • B) Inference (using the model to make predictions/generate outputs)
  • C) They consume approximately equal amounts
  • D) Neither; data storage is the primary energy cost
Answer **B)** Inference (using the model to make predictions/generate outputs) *Why B:* Training happens once or a few times, while inference happens billions of times as users interact with the system. For widely deployed systems, the cumulative energy of inference far exceeds the one-time training cost. *Why not A:* Training is more intensive per unit of time but is a one-time cost. *Why not C:* For popular systems, inference dominates significantly. *Why not D:* While storage has costs, computation (training and inference) is the dominant energy consumer. *Reference:* Section 18.1

3. Power Usage Effectiveness (PUE) is a metric that measures:

  • A) How many AI models can be trained per kilowatt-hour
  • B) The ratio of total data center energy to computing energy
  • C) The percentage of renewable energy used by a data center
  • D) The carbon emissions per unit of computation
Answer **B)** The ratio of total data center energy to computing energy *Why B:* PUE captures how efficiently a data center converts electricity into useful computation. A PUE of 1.0 would mean all energy goes to computing; higher values indicate energy lost to cooling, lighting, and other overhead. *Why not A:* PUE measures facility efficiency, not model training efficiency. *Why not C:* PUE does not measure renewable energy percentage. *Why not D:* PUE is about energy efficiency, not carbon specifically. *Reference:* Section 18.1

4. Which of the following is NOT a significant environmental cost of AI hardware beyond electricity consumption?

  • A) Water consumption for data center cooling
  • B) Mining rare earth minerals for chip manufacturing
  • C) Noise pollution from GPU fans
  • D) Toxic e-waste from discarded hardware
Answer **C)** Noise pollution from GPU fans *Why C:* While data centers can be noisy, noise pollution from individual GPU fans is not a significant global environmental concern. The chapter identifies water, minerals, and e-waste as the major non-energy environmental costs. *Why not A:* Water consumption is a significant concern, particularly in water-scarce regions. *Why not B:* Mining for rare earth elements, cobalt, lithium, and other minerals carries major environmental and human costs. *Why not D:* E-waste is one of the fastest-growing waste streams globally, with toxic materials and disproportionate impact on developing countries. *Reference:* Section 18.2

5. "Embodied carbon" in the context of AI refers to:

  • A) Carbon captured by AI-designed materials
  • B) Carbon emissions from manufacturing AI hardware, distinct from operational emissions
  • C) The amount of carbon data the model was trained on
  • D) Carbon offsets purchased by AI companies
Answer **B)** Carbon emissions from manufacturing AI hardware, distinct from operational emissions *Why B:* Embodied carbon includes emissions from mineral extraction, chip fabrication, assembly, and transportation — all occurring before the hardware is even plugged in and used. *Why not A:* This describes carbon capture technology, a separate concept. *Why not C:* This is not a meaningful concept. *Why not D:* Carbon offsets are a separate mechanism for addressing emissions. *Reference:* Section 18.2

6. AI is being used to improve climate modeling primarily by:

  • A) Replacing climate scientists entirely
  • B) Creating emulators that approximate climate simulations faster, enabling more scenarios to be tested
  • C) Reducing the carbon footprint of climate research
  • D) Generating public-facing climate change marketing materials
Answer **B)** Creating emulators that approximate climate simulations faster, enabling more scenarios to be tested *Why B:* Machine learning emulators can produce results similar to traditional physics-based climate models in a fraction of the time, allowing scientists to explore a wider range of scenarios and produce more detailed local forecasts. *Why not A:* AI augments climate scientists' capabilities; it does not replace them. *Why not C:* While efficiency matters, the primary contribution is speed and resolution, not reducing research emissions. *Why not D:* AI for climate modeling is a scientific application, not a marketing tool. *Reference:* Section 18.3

7. The rebound effect (Jevons paradox) in AI occurs when:

  • A) AI models become less efficient over time
  • B) Efficiency improvements make AI cheaper, leading to more total AI usage that offsets or exceeds the efficiency gains
  • C) Renewable energy prices increase due to AI demand
  • D) AI hardware becomes obsolete and is discarded
Answer **B)** Efficiency improvements make AI cheaper, leading to more total AI usage that offsets or exceeds the efficiency gains *Why B:* When each computation becomes more efficient, the cost per unit drops. This encourages more usage — larger models, more applications, more users — and the total environmental impact can increase even as per-unit efficiency improves. *Why not A:* The rebound effect is about the *consequences* of efficiency gains, not efficiency declines. *Why not C:* Energy price dynamics are a separate issue. *Why not D:* Hardware obsolescence contributes to e-waste but is not the rebound effect. *Reference:* Section 18.6

8. Knowledge distillation is a "green AI" technique that:

  • A) Extracts renewable energy information from training data
  • B) Trains a smaller model to approximate the performance of a larger one
  • C) Reduces the water consumption of data centers
  • D) Converts AI hardware into recyclable materials
Answer **B)** Trains a smaller model to approximate the performance of a larger one *Why B:* Knowledge distillation creates a compact "student" model that learns to mimic a larger "teacher" model, retaining most of the performance at a fraction of the computational cost for inference. *Why not A:* The technique is about model efficiency, not energy source data. *Why not C:* It reduces computational energy, not water consumption directly. *Why not D:* It is a software technique, not a hardware recycling process. *Reference:* Section 18.5

9. A company claims its AI operations are "carbon neutral." Which of the following should make you most skeptical of this claim?

  • A) The company uses modern, energy-efficient hardware
  • B) The company relies on annual average renewable energy matching rather than hourly matching, and excludes Scope 3 (supply chain) emissions
  • C) The company's data centers have a PUE of 1.15
  • D) The company has reduced its per-query energy consumption by 30%
Answer **B)** The company relies on annual average renewable energy matching rather than hourly matching, and excludes Scope 3 (supply chain) emissions *Why B:* Annual matching means the data center may run on fossil fuels during nights and winters while claiming "100% renewable" on paper. Excluding Scope 3 emissions means hardware manufacturing, mining, and e-waste are not counted. *Why not A:* Efficient hardware is a genuine positive step. *Why not C:* A PUE of 1.15 is excellent and represents real efficiency. *Why not D:* Reducing per-query energy is a legitimate improvement (though the rebound effect applies). *Reference:* Section 18.6

10. Which of the following best describes the relationship between AI and the environment?

  • A) AI is purely harmful to the environment and should be restricted
  • B) AI is a net positive for the environment because its climate applications outweigh its costs
  • C) AI is both a significant environmental burden and a powerful environmental tool, and the net effect depends on human decisions about deployment, efficiency, and energy sources
  • D) AI has no meaningful environmental impact compared to other industries
Answer **C)** AI is both a significant environmental burden and a powerful environmental tool, and the net effect depends on human decisions about deployment, efficiency, and energy sources *Why C:* This captures the chapter's central argument: the relationship is genuinely paradoxical, and the outcome depends on policy, technical, and business decisions. *Why not A:* This ignores the significant environmental benefits of AI applications. *Why not B:* This is possible but not established — and the rebound effect means it is not automatic. *Why not D:* Data center energy consumption alone represents a significant and rapidly growing share of global electricity use. *Reference:* Sections 18.1–18.7

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

11. A single query to a large language model like ChatGPT uses approximately the same energy as a traditional Google search.

Answer **False** *Explanation:* A single LLM query uses roughly 10 times the energy of a standard Google search, due to the far greater computational intensity of generating text with a large neural network compared to retrieving and ranking indexed web pages.

12. The environmental costs of AI are distributed equally across the globe.

Answer **False** *Explanation:* The benefits of AI accrue disproportionately to wealthy nations and technology companies, while environmental costs — mineral mining, e-waste processing, and climate impacts — disproportionately affect developing countries and marginalized communities. This is an environmental justice issue.

13. Making AI models more computationally efficient will automatically reduce AI's total environmental impact.

Answer **False** *Explanation:* The rebound effect (Jevons paradox) means that efficiency gains can be consumed by increased usage. If more efficient models lead to more users, more applications, and larger models, total environmental impact can increase even as per-unit efficiency improves. Policy interventions are needed to ensure efficiency gains translate into environmental gains.

14. Transfer learning — fine-tuning a pre-trained model rather than training from scratch — is an example of "green AI" in practice.

Answer **True** *Explanation:* Transfer learning can reduce training energy by a factor of 100 or more compared to training a model from scratch. It is both a standard engineering practice and a concrete example of how good technical practices can align with environmental sustainability.

15. Data center water consumption is only a concern in desert climates.

Answer **False** *Explanation:* While water scarcity amplifies the concern in arid regions, data centers consume significant water for cooling regardless of location. Even in temperate climates, the scale of consumption (millions of gallons per day for large facilities) can strain local water supplies and compete with other uses.

Section 3: Short Answer (2 points each)

16. Describe three environmental costs of AI other than carbon emissions from electricity consumption. For each, explain who is most affected.

Sample Answer (1) **Water consumption:** Data centers use millions of gallons of water daily for cooling. Communities in water-scarce regions (western U.S., Middle East) are most affected, as data centers compete with agriculture and residential water needs. (2) **Mineral mining:** AI hardware requires rare earth elements, cobalt, lithium, and other minerals. Communities near mining operations — particularly in the Democratic Republic of Congo (cobalt), China's Inner Mongolia (rare earths), and South America (lithium) — bear the environmental damage and health risks. (3) **E-waste:** Discarded AI hardware contains toxic materials and is often shipped to developing countries for informal processing. Workers and communities in these processing areas — often in West Africa, South Asia, and Southeast Asia — face direct exposure to toxins. *Rubric — full credit requires:* - Three distinct environmental costs beyond carbon/electricity - For each, a specific population or region most affected - Demonstration of understanding that costs are distributed unevenly

17. Explain the rebound effect using a specific AI example. Then describe one policy intervention that could counteract it.

Sample Answer **Example:** Suppose a new chip architecture makes AI inference 50% more energy-efficient. This lowers the cost per query, making generative AI tools affordable to deploy in more applications (customer service, education, entertainment). Companies that already use AI can now serve more queries within the same energy budget. The result: total AI energy consumption increases because the growth in usage exceeds the efficiency gain. **Policy intervention:** A carbon tax on data center emissions would ensure that even as efficiency improves, the environmental cost remains visible in the price. Companies would still benefit from efficiency gains (lower cost per unit), but the carbon price would discourage unlimited scaling and incentivize renewable energy adoption. The tax revenue could fund renewable energy infrastructure or environmental remediation. *Rubric — full credit requires:* - A clear, specific example of the rebound effect in AI - An explanation of why efficiency gains are consumed by increased use - A specific, plausible policy intervention with reasoning

18. How is AI being used to fight deforestation? Describe the technology, what it can and cannot do, and what it depends on beyond the technology itself.

Sample Answer AI systems like Global Forest Watch use machine learning to analyze satellite imagery from the Landsat constellation, detecting forest loss in near-real-time — identifying deforestation events within days of their occurrence, far faster than manual analysis could achieve. In Brazil, the DETER system provides weekly deforestation alerts that inform law enforcement operations against illegal logging. However, the technology can only *detect* deforestation — it cannot *prevent* it. The alerts are only useful if they lead to action: enforcement operations, legal consequences for violators, and policy support for forest protection. During periods when Brazil's government prioritized enforcement, DETER-informed operations significantly reduced deforestation rates. During periods of reduced political will, deforestation increased despite the monitoring technology remaining active. The technology depends on: continued satellite coverage, sufficient computing infrastructure to process imagery, and — most critically — political will to act on the information. *Rubric — full credit requires:* - Accurate description of the technology (satellite imagery + ML) - Distinction between detection and prevention - Recognition that technology depends on political and institutional context

Section 4: Applied Scenario (3–5 points)

19. A startup is developing an AI-powered agricultural optimization platform that helps farmers reduce water and fertilizer use by 20–30%, using machine learning models that process drone imagery and soil sensor data. The system requires training a computer vision model (estimated 500 GPU-hours), hosting the trained model in a cloud data center for inference, and manufacturing specialized IoT sensors for deployment in fields.

Write an environmental assessment (200–300 words) that: - Identifies the environmental costs of the system (training, inference, hardware, e-waste) - Identifies the environmental benefits (water savings, fertilizer reduction, emissions reduction) - Assesses whether the net environmental impact is likely positive, negative, or uncertain - Proposes two specific measures to improve the system's environmental profile

Sample Answer **Environmental costs:** Training at 500 GPU-hours is relatively modest (roughly 180 kWh at 300W with 1.2 PUE, producing approximately 72 kg CO2 at global average grid intensity — comparable to driving about 175 miles). Ongoing inference costs depend on the number of users but are likely modest for a specialized agricultural tool. Hardware costs include manufacturing IoT sensors (embodied carbon, mineral extraction) and their eventual disposal as e-waste. Cloud computing adds to data center energy demand. **Environmental benefits:** A 20-30% reduction in water use across participating farms could save millions of gallons annually. Reduced fertilizer use decreases nitrogen runoff (which causes algal blooms and dead zones) and reduces nitrous oxide emissions (a potent greenhouse gas). Better crop management can reduce food waste and the need to expand farmland into forests. **Net assessment:** The net environmental impact is very likely positive. The training and inference costs are modest compared to the scale of agricultural resource savings. Even a small number of farms saving 20-30% water and fertilizer would offset the system's computational carbon footprint many times over. **Improvement measures:** (1) Use transfer learning from existing agricultural vision models rather than training from scratch, reducing training costs further. (2) Design IoT sensors with modular, repairable components and establish a take-back program for end-of-life recycling, minimizing e-waste. *Rubric:* | Criterion | 0 pts | 1 pt | 2 pts | 3 pts | |-----------|-------|------|-------|-------| | Cost identification | Not addressed | Lists 1-2 costs | Covers training, inference, hardware | Includes lifecycle considerations | | Benefit identification | Not addressed | Generic mention | Specific, quantified where possible | Connects to broader environmental systems | | Net assessment | Not addressed | States positive/negative without reasoning | Reasoned assessment with evidence | Acknowledges uncertainty while making a justified call | | Improvements | Not addressed | Generic suggestion | Two specific, actionable measures | Measures address different aspects of the footprint |

20. You are advising a university that wants to reduce the environmental impact of its AI research without compromising research quality. The university currently has a GPU cluster that runs 24/7, researchers train models without energy budgets, and there is no requirement to report computational costs in publications. Draft a set of five specific policy recommendations (150–250 words total) with brief justifications for each.

Sample Answer **1. Implement computational budgets.** Require research groups to estimate and justify the GPU-hours needed for each project during the proposal stage. This does not prevent large experiments but ensures they are intentional, not wasteful. **2. Require carbon reporting in publications.** Mandate that all papers from the university report estimated energy consumption and CO2 emissions for their experiments, using tools like CodeCarbon or ML CO2 Impact Calculator. This normalizes transparency and incentivizes efficiency. **3. Schedule GPU-intensive work during low-carbon grid periods.** Electricity grids are cleaner at some times of day (when solar and wind are producing) than others. Schedule large training runs to coincide with high-renewable periods when possible. **4. Incentivize efficiency.** Create an annual "Efficient AI" award for the research group that achieves the best performance per computational unit. Celebrate researchers who find clever ways to achieve strong results with fewer resources. **5. Invest in green infrastructure.** Purchase renewable energy for the GPU cluster (either on-site solar or verified renewable energy contracts). Establish an e-waste recycling program for retired hardware. Consider cooling system upgrades to reduce water consumption. *Rubric:* | Criterion | 0 pts | 1 pt | 2 pts | 3 pts | |-----------|-------|------|-------|-------| | Specificity | Generic advice | Some specific proposals | Five distinct, actionable policies | Policies address multiple dimensions (energy, reporting, incentives, infrastructure) | | Feasibility | Impractical | Some practical | Most realistic for a university setting | Considers costs and implementation challenges |

Scoring & Next Steps

Score Assessment Recommended Action
< 50% Needs review Re-read sections 18.1–18.3, redo Part A exercises
50–70% Partial Review weak areas, redo Part B exercises
70–85% Solid Ready to proceed; revisit any missed topics
> 85% Strong Proceed; consider Deep Dive extensions