AI, Society, Work, and Environmental Sustainability

Introduction

Artificial intelligence is not only a technical capability. It is also a social, economic, political, cultural, and environmental force. AI systems influence who is hired, monitored, educated, insured, represented, or excluded. They can improve accessibility and public services, yet they can also intensify inequality, displace workers, consume substantial energy, and amplify misinformation. Responsible innovation therefore requires more than building accurate models. It requires examining how systems are embedded in institutions, markets, workplaces, communities, and ecological systems.

This chapter examines the relationships between AI, work, society, democracy, culture, and environmental sustainability. It develops a practical approach for assessing impacts across the complete AI lifecycle: data collection, model development, deployment, use, maintenance, and retirement. The central argument is that responsible AI must be socially beneficial, inclusive, accountable, and environmentally proportionate. Technical performance is necessary but insufficient.

Learning Objectives

By the end of this chapter, students should be able to:

1. Explain how AI can transform employment, job quality, skills, and workplace power.

2. Distinguish labor displacement from task transformation and augmentation.

3. Evaluate workplace monitoring and automated management systems ethically.

4. Analyze accessibility, digital inequality, and cultural impacts of AI.

5. Identify threats AI may pose to democratic integrity and public reasoning.

6. Assess energy use, carbon emissions, water use, and material consumption across the AI lifecycle.

7. Apply a just-transition framework to organizational AI adoption.

8. Propose sustainable-AI strategies and governance controls.

1. AI and the Transformation of Work

1.1 Jobs, tasks, and occupations

Predictions that AI will either eliminate all jobs or create unlimited prosperity are both overly simplistic. Most occupations consist of multiple tasks, and AI may automate some tasks while supporting or increasing demand for others. A useful distinction is:

• Labor displacement: workers lose tasks, hours, or employment because an AI system performs activities previously performed by people.
• Task augmentation: AI assists workers while human responsibility remains significant.
• Job transformation: the content, pace, skills, supervision, or quality of a job changes substantially.
• Job creation: new roles emerge, such as AI auditors, data stewards, model-risk specialists, and human-in-the-loop supervisors.

The consequences depend on technical capability, implementation cost, regulation, consumer demand, organizational strategy, and worker bargaining power. An AI tool that can technically perform a task may not replace workers if errors are costly, customers value human interaction, or laws require human judgment. Conversely, even partial automation can reduce staffing when an organization redesigns workflows around the system.

1.2 Substitution and complementarity

AI tends to substitute for routine, predictable, and codifiable activities. It can complement workers when it provides information, reduces administrative effort, or supports complex decisions without removing human discretion. However, complementarity is not automatically beneficial. A system may increase productivity while also increasing workload, reducing autonomy, or transferring responsibility to workers who cannot challenge the technology.

The distribution of gains is a central ethical issue. If productivity gains flow mainly to owners and senior managers while workers experience reduced wages, insecurity, or intensified surveillance, innovation may be technically successful but socially unjust. Responsible innovation asks who benefits, who bears risks, and who participates in decisions.

1.3 The quality of work

Employment should not be evaluated only by the number of jobs. Job quality includes pay, stability, autonomy, safety, dignity, learning opportunities, social interaction, and access to grievance mechanisms. Algorithmic management can allocate shifts, rate performance, assign tasks, and determine pay. Such systems may improve coordination, but opaque scoring can make workers unable to understand or contest decisions.

A warehouse worker, for example, may be evaluated against algorithmically generated productivity targets that ignore equipment failure, disability, training responsibilities, or unsafe conditions. A delivery worker may be penalized by a routing system that does not account for weather or inaccessible buildings. These are not merely accuracy problems; they concern power, due process, and the meaning of fair work.

2. Workplace Monitoring and Algorithmic Management

Workplace monitoring includes the collection and analysis of attendance, location, keystrokes, communications, biometric signals, productivity indicators, and behavioral data. AI may infer fatigue, engagement, emotional state, or the likelihood of resignation. Such inferences are often uncertain and may be especially intrusive.

A responsible monitoring program should satisfy the following conditions:

1. Purpose limitation: collect data for a specific legitimate purpose rather than general surveillance.

2. Necessity and proportionality: use the least intrusive method that can achieve the purpose.

3. Transparency: explain what is collected, how it is analyzed, and how decisions are made.

4. Accuracy and contestability: provide correction procedures and a meaningful appeal process.

5. Access control: restrict sensitive information and define retention periods.

6. Non-discrimination: test outcomes across relevant worker groups.

7. Human accountability: identify decision-makers who remain responsible for consequences.

8. Worker participation: consult employees, unions, and representatives before deployment.

A key distinction is between monitoring outputs and employment decisions. An alert that a machine may require maintenance is different from a score that affects promotion or dismissal. High-impact uses require stronger validation, documentation, human review, and procedural safeguards.

Emotion-recognition systems illustrate the problem of questionable validity. Facial movements, voice patterns, or physiological signals do not reliably reveal a person’s internal emotional state across cultures and contexts. Using such systems to assess honesty, commitment, or suitability can create discrimination while appearing scientifically objective.

3. Accessibility and Digital Inequality

3.1 Accessibility as a design obligation

Accessibility means that people with different physical, sensory, cognitive, linguistic, and technological characteristics can use a system effectively. AI can improve accessibility through speech recognition, captioning, image description, translation, predictive text, navigation assistance, and personalized interfaces. These tools can increase independence and participation.

However, accessibility features can fail unevenly. Speech-recognition systems may perform worse for regional accents or people with speech impairments. Image-description tools may omit culturally important information. Automated captions may misrepresent technical terms or names. A system is not accessible simply because it includes an accessibility feature; performance must be evaluated with the communities who depend on it.

Good practice includes compliance with applicable accessibility standards, multimodal interaction, adjustable settings, compatibility with assistive technologies, user testing with disabled participants, and non-AI alternatives when automation fails. Accessibility should be considered from requirements analysis onward, not added after deployment.

3.2 The digital divide

Digital inequality involves unequal access to devices, connectivity, electricity, technical support, digital skills, relevant content, and the ability to benefit from technology. It is often described as a divide between those online and offline, but inequality is multidimensional. A person may have internet access yet lack affordable high-quality service, privacy, literacy, or confidence.

AI can deepen inequality when essential services become accessible mainly through expensive devices, proprietary platforms, or high-bandwidth interfaces. Small businesses, rural communities, minority-language speakers, and low-income schools may lack the infrastructure needed to adopt advanced tools. Data scarcity can also produce poor services for groups that are underrepresented in training data.

Inclusive strategies include public connectivity, affordable devices, local-language systems, open educational resources, community technology centers, accessible design, procurement requirements, and support for small organizations. Impact assessment should ask not only whether an average user benefits, but whether vulnerable groups are excluded or placed at additional risk.

4. Democracy, Information, and Cultural Life

4.1 Democratic integrity

Democratic systems depend on informed citizens, trustworthy institutions, pluralistic debate, and the ability to identify persuasion. Generative AI can produce convincing text, images, audio, and video at low cost. It can support political communication and civic education, but it can also enable impersonation, fabricated evidence, targeted manipulation, voter suppression, and large-scale harassment.

The danger is not only that people believe false content. Synthetic media can also create uncertainty about genuine evidence, producing a “liar’s dividend” in which real recordings are dismissed as fake. Recommendation systems may amplify sensational or polarizing material because engagement is commercially valuable. Automated political targeting can exploit personal vulnerabilities while making campaign messages difficult for the public to inspect.

Safeguards include provenance systems, labeling of synthetic content where feasible, political-ad transparency, limits on microtargeting, rapid response to coordinated manipulation, independent auditing, media literacy, and protection for journalists and researchers. Provenance tools are useful but not perfect: metadata can be removed, and authentic content may be altered without detection. Governance must therefore combine technical, legal, institutional, and educational measures.

4.2 Cultural impacts and epistemic diversity

AI models reflect the data, languages, values, and institutions from which they are developed. Systems trained predominantly on powerful languages and dominant cultural sources may marginalize local knowledge, dialects, oral traditions, and minority perspectives. Generative systems may flatten cultural differences into stereotyped representations or reproduce the styles of artists without appropriate consent or compensation.

Cultural responsibility involves representation, consent, attribution, benefit sharing, and community governance. Indigenous data governance, for example, may require collective authority over how cultural knowledge is collected and used. A dataset can be legally obtainable yet ethically inappropriate if it contains sacred, sensitive, or communally controlled material.

Organizations should conduct cultural-impact assessments, involve relevant communities, support local-language development, document data provenance, and create opt-out or licensing mechanisms where appropriate. Cultural diversity should be treated as a source of knowledge and resilience, not merely as a market feature.

5. Environmental Sustainability and AI

5.1 The AI lifecycle

Environmental impacts occur throughout an AI system’s lifecycle:

1. Extraction of minerals and production of chips, servers, and networking equipment.

2. Construction and operation of data centers.

3. Data collection, storage, preprocessing, and model training.

4. Model deployment, inference, and user-device operation.

5. Hardware replacement, recycling, and disposal.

Training can be energy-intensive, but repeated inference may dominate total energy use for a widely deployed system. A lifecycle assessment should therefore avoid focusing only on the one-time training event.

5.2 Energy, emissions, and water

Energy consumption is commonly estimated as:

\[

E = P \times t

\]

where \(E\) is energy, \(P\) is average power, and \(t\) is operating time. If power is measured in kilowatts and time in hours, energy is measured in kilowatt-hours. This relationship assumes power is represented as an appropriate average over the period.

Operational carbon emissions can be estimated as:

\[

C = E \times I

\]

where \(C\) is carbon dioxide-equivalent emissions, \(E\) is energy use, and \(I\) is the carbon intensity of electricity, usually expressed in kilograms of CO₂e per kilowatt-hour. The estimate is only as reliable as the boundaries, electricity data, and assumptions used. Location, time of use, renewable-energy claims, embodied emissions, and data-center efficiency can materially change results.

Water may be consumed directly for cooling and indirectly through electricity generation and semiconductor manufacturing. Hardware also has embodied environmental impacts from mining, manufacturing, transport, and disposal. Therefore, “using renewable electricity” does not eliminate all environmental costs.

5.3 Efficiency is not sufficient

Efficient models can reduce energy per prediction, but lower operating costs may increase total use—a rebound effect. For example, an organization may deploy a smaller model in thousands of additional applications. Sustainable AI therefore requires both efficiency and demand governance.

Useful strategies include:

• selecting the smallest model that meets the task requirement;
• reducing unnecessary retraining and duplicate experimentation;
• using efficient architectures, quantization, pruning, and caching where quality permits;
• scheduling flexible workloads when lower-carbon electricity is available;
• improving server utilization and cooling efficiency;
• measuring energy, emissions, water, and hardware impacts;
• extending hardware life and responsibly recycling equipment;
• preferring retrieval or rules-based methods when generative AI is unnecessary;
• including environmental criteria in procurement and model cards.

Efficiency methods can create trade-offs. Compression may reduce accuracy for minority-language or safety-critical cases. A carbon-saving location may increase latency or water stress. Decisions should therefore consider social as well as environmental consequences.

6. Frameworks for Responsible and Just Innovation

6.1 The just-transition perspective

A just transition asks how societies can adopt technological and environmental change without unfairly imposing costs on workers and communities. In AI adoption, this means combining innovation with participation, protection, and opportunity.

A practical sequence is:

Anticipate: identify affected tasks, occupations, groups, communities, and environmental resources before deployment.

Engage: involve workers, users, affected communities, accessibility experts, and domain professionals in system design and procurement.

Protect: provide privacy safeguards, fair procedures, income and employment protections, health and safety controls, and alternatives when systems fail.

Prepare: fund training, paid transition time, digital literacy, and pathways to genuinely available jobs rather than generic courses.

Share: distribute productivity gains through wages, reduced working time, public benefits, service improvements, or community investment.

Remedy: establish complaint, appeal, correction, compensation, and system-shutdown mechanisms.

This approach recognizes that retraining alone cannot solve displacement. Workers may lack time, financial security, geographic mobility, or available suitable jobs. Transition planning must address institutional and economic conditions.

6.2 A multidimensional impact assessment

Before deployment, an organization can assess a proposed AI system across six dimensions:

1. Work: Which tasks change? Who may lose income, autonomy, or safety?

2. Rights: Are privacy, equality, expression, due process, and human dignity affected?

3. Inclusion: Can disabled, low-income, rural, linguistic, and culturally diverse groups use and benefit from it?

4. Democracy and culture: Could it distort public reasoning, representation, or community knowledge?

5. Environment: What are energy, carbon, water, material, and end-of-life impacts?

6. Accountability: Who can explain, challenge, correct, or stop the system?

The assessment should identify alternatives, not merely evaluate a predetermined system. Sometimes the responsible decision is not to automate. Documentation should record assumptions, affected stakeholders, residual risks, mitigation owners, monitoring indicators, and review dates.

7. Worked Example: Estimating Operational Emissions

An organization operates an AI service using an average IT load of 80 kilowatts for 10 hours per day over 30 days. Assume, for illustration, that the stated power is the relevant average power and that electricity has a carbon intensity of 0.40 kilograms CO₂e per kilowatt-hour.

First calculate energy:

\[

E = P \times t = 80\text{ kW} \times (10 \times 30)\text{ h} = 24{,}000\text{ kWh}.

\]

Then calculate operational emissions:

\[

C = E \times I = 24{,}000\text{ kWh} \times 0.40\text{ kg CO₂e/kWh} = 9{,}600\text{ kg CO₂e}.

\]

Thus, the estimated operational emissions are 9,600 kg CO₂e, or 9.6 metric tonnes CO₂e, under the stated assumptions. This excludes embodied emissions from hardware, water impacts, and emissions associated with upstream infrastructure. If the service is moved to a grid with a different carbon intensity, the result changes even if computing activity remains constant. The example demonstrates why organizations should disclose boundaries and assumptions rather than presenting a single number as universally precise.

8. Applied Case: AI Scheduling in a Public Hospital

Background

The following is an illustrative case. A public hospital introduces an AI scheduling system to predict patient demand and allocate nurses across shifts. The system uses historical admissions, department workload, staff availability, and absence data.

Problem

The hospital expects shorter waiting times and better use of staff. Nurses, however, report that the system produces unstable schedules, increases short-notice changes, and penalizes staff who request accommodations. Historical data also reflect chronic understaffing in some departments, so the model interprets delayed care as normal demand rather than as evidence of resource shortages. Staff representatives are not given access to meaningful explanations.

Analysis

The project creates several risks. First, optimization of staffing efficiency may conflict with worker wellbeing and patient safety. Second, historical data encode organizational inequities. Third, accommodation-related information is sensitive and may expose workers to discrimination. Fourth, a technically accurate demand forecast does not justify automated decisions about individuals. Finally, the environmental cost of continuous prediction and optimization should be proportionate to the public benefit.

Approach and solution

The hospital establishes a joint governance group including nurses, clinicians, scheduling staff, disability representatives, data-protection specialists, and management. It changes the system’s role from automatic scheduling to decision support. Human schedulers may reject recommendations and must record reasons for significant deviations.

The hospital removes unnecessary sensitive variables, separates accommodation management from performance assessment, and tests schedules for unequal effects across departments and worker groups. It introduces minimum notice periods, limits consecutive high-intensity shifts, and creates a rapid appeal process. The model is evaluated not only on forecasting accuracy but also on patient waiting time, schedule stability, overtime, sick leave, staff perceptions, and safety incidents.

The hospital also compares the environmental cost of frequent retraining with less computationally intensive forecasting methods. It publishes a plain-language system description and conducts periodic review with staff representatives.

Outcome

In this illustrative case, the system improves demand visibility but is not allowed to determine schedules without human judgment. Some predicted efficiency gains are sacrificed to preserve stability, safety, and fairness. The result demonstrates that responsible innovation may involve redesigning objectives and constraints rather than simply improving model accuracy.

Learning points

• Historical data are records of institutional practice, not neutral truth.
• Human oversight is meaningful only when people have authority, time, and information to intervene.
• Worker participation can reveal risks invisible to developers.
• Success metrics should include rights, wellbeing, equity, safety, and environmental costs.
• A less automated system may be more responsible than a technically optimal one.

9. Governance, Regulation, and Professional Practice

Organizations should assign clear accountability for social and environmental effects. Relevant controls include impact assessments, data-protection reviews, accessibility testing, algorithmic audits, labor consultation, environmental reporting, incident management, procurement standards, and independent review for high-impact applications.

Legal obligations differ across jurisdictions, but common regulatory themes include privacy, discrimination, employment rights, product safety, consumer protection, accessibility, environmental disclosure, and transparency. Compliance is a minimum requirement, not a complete ethical justification. A system may comply with law while producing unacceptable cultural harm or excessive resource use.

Professionals should maintain documentation of data sources, model purpose, limitations, evaluation populations, energy assumptions, known failure modes, human responsibilities, and retirement criteria. Contracts with vendors should prohibit unapproved secondary use, permit audits, specify incident notification, and define responsibility for failures. Organizations should also plan for system decommissioning, data deletion, hardware reuse, and preservation of essential human expertise.

10. Challenges and Limitations

Impact assessment is difficult because effects may emerge gradually and interact with broader economic forces. It is also difficult to attribute social outcomes to one AI system. Stakeholders may disagree about what counts as harm or benefit. Environmental estimates can be uncertain because vendors may not disclose energy, hardware, or water data. Efficiency benchmarks can privilege average performance and hide failures for minority groups.

There are also tensions among objectives. More transparency may expose security-sensitive information. More human review may increase cost and delay. Local processing may improve privacy but require additional hardware. A highly accessible interface may demand more computation. These tensions do not justify ignoring impacts; they require explicit trade-offs, stakeholder participation, and documented reasons.

A further limitation is automation bias: users may trust an AI recommendation because it appears objective. Training and interface design should make uncertainty visible and encourage critical review. Conversely, excessive warnings can lead to alert fatigue. Responsible systems must calibrate both automation and human responsibility.

Contemporary Developments and Best Practices

Current developments include foundation models, multimodal systems, AI agents, edge computing, synthetic data, and increasingly automated workplace platforms. These developments create opportunities for accessibility and productivity but also increase concerns about scale, concentration of power, copyright, labor conditions in data production, and energy demand.

Best practices for students and professionals include:

• define the social purpose before selecting a model;
• assess whether AI is necessary;
• map affected stakeholders and unequal impacts;
• involve workers and communities early;
• test performance across relevant populations and contexts;
• provide accessible, understandable explanations;
• preserve human appeal and correction mechanisms;
• measure lifecycle environmental impacts;
• disclose uncertainty and assumptions;
• monitor real-world outcomes after deployment;
• distribute benefits and support just transitions;
• stop or redesign systems whose harms cannot be controlled.

Chapter Summary

AI changes work by automating tasks, augmenting workers, transforming job quality, and creating new roles. The ethical question is not simply how many jobs exist, but who controls technology, who benefits from productivity, and whether workers retain dignity, safety, autonomy, and recourse. Workplace monitoring must be necessary, proportionate, transparent, secure, contestable, and subject to worker participation.

Accessibility requires inclusive design and evaluation with disabled and diverse users. Digital inequality includes differences in connectivity, affordability, skills, language, privacy, and institutional capacity. AI can strengthen or undermine democracy through synthetic media, recommendation systems, and targeted persuasion. Cultural responsibility requires representation, consent, attribution, community governance, and support for minority knowledge and languages.

Environmental sustainability requires lifecycle thinking. Energy use can be estimated as power multiplied by time, while operational emissions depend on energy and carbon intensity; these estimates must state boundaries and assumptions. Sustainable AI combines efficient models, responsible demand, clean and appropriately located infrastructure, hardware stewardship, and transparent measurement. A just transition anticipates impacts, engages stakeholders, protects rights, prepares workers, shares gains, and provides remedies. Responsible innovation is therefore a continuing governance practice rather than a one-time technical approval.

Key Terms

Labor displacement: Reduction or loss of human work caused by automation or organizational redesign.

Task augmentation: Use of AI to assist human performance without fully replacing human responsibility.

Job transformation: Significant change in a job’s tasks, skills, autonomy, or conditions.

Algorithmic management: Use of computational systems to allocate, monitor, evaluate, or direct work.

Workplace surveillance: Collection and analysis of worker information for monitoring or control.

Contestability: The ability to question, appeal, correct, or challenge an automated decision.

Accessibility: Design that enables people with diverse abilities to use a system effectively.

Digital inequality: Unequal access to technology, connectivity, skills, benefits, or influence.

Democratic integrity: Conditions supporting informed participation, trustworthy elections, and pluralistic debate.

Synthetic media: AI-generated or AI-manipulated text, image, audio, or video.

Cultural impact: Effects of technology on identities, traditions, representation, language, and knowledge.

Lifecycle assessment: Evaluation of environmental impacts across production, use, and disposal.

Carbon intensity: Emissions associated with a unit of electricity, commonly kg CO₂e/kWh.

Embodied emissions: Emissions associated with materials, manufacturing, transport, and infrastructure.

Rebound effect: Increased total consumption resulting from efficiency improvements.

Just transition: Fair management of technological or environmental change for affected workers and communities.

Human oversight: Human capacity and authority to review, intervene in, and stop AI decisions.

Sustainable AI: AI designed and governed to reduce environmental harm while preserving social benefit.

Review Questions

Short Answer Questions

1. Distinguish labor displacement, task augmentation, and job transformation.

2. State four principles for responsible workplace monitoring.

3. Why is accessibility testing with affected communities important?

4. What is the difference between operational and embodied environmental emissions?

5. Identify three elements of a just transition.

Long Answer Questions

1. Explain how AI can improve productivity while reducing job quality. Use workplace examples.

2. Discuss how AI-generated media and recommendation systems can affect democratic integrity.

3. Explain why digital inequality is multidimensional and propose institutional responses.

4. Describe a lifecycle approach to assessing the environmental impact of an AI system.

5. Evaluate the role of worker participation in responsible algorithmic management.

Analytical / Application Questions

1. An employer proposes using keystroke monitoring and emotion recognition to identify “unproductive” employees. Analyze the proposal using necessity, proportionality, validity, privacy, discrimination, and contestability criteria. Recommend an alternative.

2. A university plans to deploy a generative AI tutor accessible only through a high-bandwidth mobile application. Develop an assessment covering accessibility, digital inequality, cultural inclusion, privacy, learning quality, and environmental sustainability.

3. Using the worked-example assumptions in this chapter, explain how operational emissions would change if the system ran for 20 hours per day instead of 10, and identify at least three impacts not captured by the calculation.

Multiple Choice Questions with Answers

Question 1

Which situation best illustrates task augmentation?

A. An AI system completely eliminates a customer-service department.

B. An AI assistant summarizes cases while a professional reviews and decides.

C. A company dismisses workers after installing a fully automated line.

D. A new law prohibits a particular occupation.

Correct Answer: B. An AI assistant summarizes cases while a professional reviews and decides.

Question 2

Which principle requires collecting only data needed for a defined workplace purpose?

A. Purpose limitation

B. Model scaling

C. Open-ended surveillance

D. Data maximization

Correct Answer: A. Purpose limitation

Question 3

Why can historical workplace data create unfair AI outcomes?

A. Historical data are always too small to analyze.

B. Historical data may encode existing institutional inequalities.

C. Historical data cannot contain numerical variables.

D. Historical data automatically remove human bias.

Correct Answer: B. Historical data may encode existing institutional inequalities.

Question 4

What is a central concern about AI-generated political media?

A. It always has lower production quality than human media.

B. It can enable impersonation and make genuine evidence easier to deny.

C. It cannot be distributed through online platforms.

D. It is legally identical in every jurisdiction.

Correct Answer: B. It can enable impersonation and make genuine evidence easier to deny.

Question 5

If an AI service uses 50 kW for 4 hours, what energy does it consume under the average-power assumption?

A. 12.5 kWh

B. 46 kWh

C. 200 kWh

D. 2,000 kWh

Correct Answer: C. 200 kWh

Question 6

Using carbon intensity, operational emissions are estimated by multiplying:

A. Power by model accuracy

B. Energy use by carbon intensity

C. Water use by working hours

D. Hardware mass by latency

Correct Answer: B. Energy use by carbon intensity

Question 7

Which statement best describes the rebound effect?

A. Efficiency reduces both energy per task and total demand in every case.

B. Efficiency improvements may lower cost and encourage greater total use.

C. Renewable electricity eliminates embodied emissions.

D. Model compression always harms accessibility.

Correct Answer: B. Efficiency improvements may lower cost and encourage greater total use.

Question 8

Which action is most consistent with a just transition?

A. Deploying automation without consulting workers

B. Providing generic training after dismissals occur

C. Involving workers, protecting affected people, and sharing productivity gains

D. Measuring only model accuracy

Correct Answer: C. Involving workers, protecting affected people, and sharing productivity gains

Question 9

Why is technical accuracy insufficient for evaluating an AI scheduling system?

A. Accuracy cannot be calculated for any AI system.

B. Scheduling may also affect safety, fairness, workload, and worker autonomy.

C. Technical accuracy is unrelated to predictions.

D. Human review makes accuracy irrelevant.

Correct Answer: B. Scheduling may also affect safety, fairness, workload, and worker autonomy.

Question 10

Which is an example of embodied environmental impact?

A. Electricity used during one prediction

B. Emissions from manufacturing and transporting servers

C. Time taken by a user to read an output

D. The model’s classification accuracy

Correct Answer: B. Emissions from manufacturing and transporting servers