Ethics, Governance, Standards and Sustainability

Introduction and Context

The Internet of Things (IoT) connects sensors, actuators, vehicles, appliances, industrial equipment, buildings, medical devices and public infrastructure to networks and computational services. Artificial intelligence (AI) adds capabilities such as prediction, classification, anomaly detection, optimisation and automated decision-making. Together, IoT and AI can improve healthcare, energy management, transport, manufacturing, agriculture and public services. However, the same systems can also create serious risks: continuous surveillance, discriminatory decisions, unsafe automation, insecure data practices, environmental damage and unclear responsibility when something goes wrong.

Ethics and governance are therefore not optional additions to technical design. They determine what data may be collected, for which purposes, by whom, under what safeguards and with what consequences for people and the environment. Responsible IoT and AI requires a lifecycle perspective covering design, procurement, deployment, operation, maintenance, decommissioning and disposal.

This chapter examines ethical principles, governance mechanisms, standards, auditability, compliance and sustainability. It also considers how engineers and managers can translate broad principles into requirements, controls, evidence and operational decisions.

Learning Objectives

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

1. Explain major ethical risks in IoT and AI systems.

2. Analyse algorithmic bias, surveillance, consent, transparency and human agency.

3. Design governance arrangements for data, models, devices and suppliers.

4. Describe auditability, accountability and compliance mechanisms.

5. Apply standards and risk-management frameworks to practical situations.

6. Evaluate energy use, e-waste and lifecycle sustainability.

7. Recommend responsible innovation practices for an IoT–AI project.

1. Ethical Foundations for Connected Intelligence

Ethics concerns justified judgments about what ought to be done. In IoT and AI, ethical analysis supplements legal compliance. An action may be legal yet still be unfair, intrusive or socially harmful. Conversely, ethical expectations may develop before legislation becomes specific.

Several principles are especially relevant:

Respect for autonomy requires that people are not manipulated or subjected to significant decisions without meaningful choice. Beneficence requires seeking beneficial outcomes, while non-maleficence requires avoiding foreseeable harm. Justice and fairness require that benefits and burdens are distributed appropriately and that comparable people are not treated arbitrarily differently. Privacy protects individuals’ control over information and freedom from unjustified observation. Accountability requires identifiable responsibility for decisions and impacts. Transparency requires sufficient information for appropriate understanding, challenge and oversight. Sustainability extends concern to future generations, natural resources and ecological systems.

These principles can conflict. For example, collecting detailed location data may improve emergency response but increase privacy risks. A highly accurate facial-recognition system may still be unacceptable in a public setting if it enables disproportionate surveillance. Ethical engineering therefore involves identifying affected parties, assessing competing values, documenting trade-offs and selecting safeguards.

Stakeholders and impact assessment

Stakeholders include direct users, people observed by sensors, workers affected by automation, customers, communities, operators, suppliers, regulators and future users. Some are not represented in procurement decisions; children, people with disabilities and marginalised groups may face particular risks.

A useful impact assessment asks:

• What is the intended purpose and what secondary uses are possible?
• Whose data is collected, inferred or shared?
• Who benefits and who bears risk?
• What errors are likely, and are their consequences reversible?
• Can individuals understand, contest or opt out of the system?
• What happens when connectivity, sensors or models fail?
• What environmental resources are consumed throughout the lifecycle?

The assessment should be repeated when the system, data, model, supplier or operating context changes.

2. Algorithmic Bias and Fairness

Algorithmic bias is a systematic pattern of error or disadvantage associated with data, system design, implementation or use. Bias does not necessarily require discriminatory intent. It can arise from historical inequality, unrepresentative samples, faulty sensors, proxy variables, labelling practices, objective functions or unequal access to appeal mechanisms.

Sources of bias in IoT–AI systems

1. Representation bias: sensor data may cover some populations or locations more thoroughly than others.

2. Measurement bias: devices may measure groups differently because of calibration, skin tone, language, clothing, connectivity or environmental conditions.

3. Historical bias: training data may reflect past discriminatory decisions.

4. Aggregation bias: one model is applied to groups with different patterns or needs.

5. Label bias: labels may encode subjective or institutionally biased judgments.

6. Deployment bias: a model developed for one context is used in another.

7. Feedback loops: predictions influence behaviour, and resulting data reinforces the original prediction.

Fairness is not a single universal metric. Possible approaches include equalising error rates, examining selection rates, improving calibration or ensuring procedural fairness. These goals can conflict, particularly when groups have different base rates. Consequently, fairness must be defined in relation to the decision, affected rights, institutional purpose and legal context rather than selected solely because a metric is convenient.

A practical fairness assessment should disaggregate performance by relevant groups, examine false positives and false negatives, test data quality, consult affected communities and monitor outcomes after deployment. Technical parity does not prove social justice: a system can have similar error rates while being used for an unjust purpose.

Mitigation strategies

Mitigation may occur before modelling, during modelling or after deployment. Examples include improving sampling, recalibrating sensors, removing unjustified proxies, reweighting training data, using fairness constraints, setting different operating thresholds only where justified, adding human review and providing appeal procedures. Documentation should explain why a mitigation was selected and what trade-offs it creates.

3. Surveillance, Privacy and Consent

IoT devices often collect data continuously, passively and at a scale that people cannot easily perceive. Smart cameras, microphones, wearables, location trackers and connected workplaces may reveal routines, relationships, health conditions and political or religious activities. AI can infer sensitive attributes even when those attributes were not directly collected.

Privacy has several dimensions: confidentiality, control over personal information, anonymity, freedom from persistent monitoring and protection against harmful inference. Security is necessary for privacy but is not identical to it. A securely stored dataset can still have been collected for an unjustified purpose.

Data minimisation and purpose limitation

Responsible systems collect only data necessary for a specified and legitimate purpose, retain it for no longer than needed and restrict secondary use. Designers should consider processing data locally, extracting features rather than transmitting raw streams, reducing sampling frequency and deleting information at source. Anonymisation may reduce risk but is not automatically permanent; combining datasets can enable re-identification. Pseudonymisation protects identity through separation of identifiers, but the data remains personal when re-identification is possible.

Meaningful consent

Consent should be informed, specific, freely given and revocable where applicable. Long notices, preselected options and consent bundled with essential services often fail to provide meaningful choice. In public spaces, consent may be impractical or impossible; signage alone does not necessarily justify pervasive monitoring. Alternatives may include legal authority, strong public-interest justification, strict necessity tests, opt-out mechanisms, short retention and independent oversight. Special care is required for children, employees and people with limited bargaining power.

4. Transparency, Explainability and Human Agency

Transparency means making relevant information available. It does not mean publishing every line of source code. Different audiences need different information:

• Individuals need to know that automated processing occurs, its purpose, significant consequences and how to challenge it.
• Operators need system status, uncertainty, limitations and escalation procedures.
• Auditors need data lineage, model versions, testing evidence and access records.
• Managers need risk, performance, cost and accountability information.
• Regulators may require documentation, testing results and incident reports.

Explainability is the ability to provide understandable reasons or representations for a system’s output. A feature-importance chart may describe association but not establish causation. A post-hoc explanation may be plausible without faithfully representing the model. Explanations should therefore be tested for fidelity, stability and usefulness to the intended audience.

Human-in-the-loop systems require human input during decision-making. Human-on-the-loop systems allow people to supervise automated operation and intervene. Human-in-command arrangements reserve authority to determine objectives, limits and suspension of the system. Human involvement is meaningful only when the person has competence, time, authority, relevant information and freedom to reject the recommendation. Merely placing a person near an automated system can create automation bias, in which the human accepts outputs uncritically.

High-impact or safety-critical decisions should have defined override conditions, fail-safe states, escalation paths and appeal mechanisms. Human agency also includes the ability to refuse unnecessary monitoring and to understand how one’s data affects opportunities or services.

5. Accountability and Responsibility

Accountability means that an organisation can explain and justify its actions, demonstrate compliance and provide remedy. Responsibility should be allocated before deployment, not improvised after an incident.

A typical responsibility structure includes:

• System owner: accountable for purpose, resources and risk acceptance.
• Product or service team: responsible for requirements, testing and maintenance.
• Data steward: responsible for provenance, quality, access and retention.
• Model owner: responsible for model development, validation and monitoring.
• Security and privacy teams: responsible for threat and privacy controls.
• Operations staff: responsible for safe configuration and incident response.
• Supplier: responsible for contracted deliverables, disclosures and support.
• Independent assurance function: challenges evidence and reports deficiencies.
• Senior governance body: approves high-risk use and accepts residual risk.

A responsibility matrix such as RACI—Responsible, Accountable, Consulted and Informed—can clarify roles, but it does not replace substantive accountability. Contracts should address data rights, security updates, model changes, audit access, incident notification, subcontractors, service termination and deletion or return of data.

6. Data and AI Governance

Governance is the system of policies, roles, controls, decision rights and oversight through which an organisation directs technology. Effective governance covers the entire data and AI lifecycle.

Data governance

Core data-governance controls include:

• inventory and classification of datasets;
• documented provenance and lawful basis for collection;
• quality checks and known limitations;
• access controls based on need;
• retention and deletion schedules;
• data-sharing approvals;
• lineage from collection to output;
• procedures for correction and individual requests;
• security monitoring and breach response.

IoT data requires additional attention to device identity, timestamp integrity, sensor calibration, missing values, firmware changes and edge-to-cloud transmission. Data quality is not merely a technical property; a dataset may be accurate but irrelevant, or complete but collected without legitimate authority.

AI lifecycle governance

A governed AI lifecycle commonly includes:

1. Use-case approval: define purpose, affected people, risk level and prohibited uses.

2. Design review: evaluate necessity, alternatives, privacy, fairness, safety and sustainability.

3. Data review: document provenance, consent or authority, quality and representativeness.

4. Development controls: version code, data and models; separate development and production.

5. Validation: test performance, robustness, security, fairness and uncertainty.

6. Deployment approval: define operating limits, human roles and rollback conditions.

7. Continuous monitoring: track drift, incidents, complaints, energy use and group outcomes.

8. Change management: reassess risks when models, sensors, data or context change.

9. Retirement: revoke access, preserve necessary records and securely delete or recycle components.

Risk classification should consider not only model accuracy but also the scale, reversibility and distribution of harm. A low-accuracy recommendation system may be low risk, whereas a highly accurate system used to deny essential services may be high risk.

7. Auditability, Documentation and Compliance

Auditability is the ability to reconstruct what a system did, with which data, model, configuration and authority, and why relevant controls were considered adequate. Auditability supports accountability, debugging, regulatory review and learning from incidents.

Useful evidence includes:

• system and data-flow diagrams;
• asset and model inventories;
• data sheets and model cards;
• risk and impact assessments;
• consent and access records;
• version-control histories;
• test protocols and results;
• approval decisions;
• logs of inputs, outputs, overrides and alerts;
• incident and complaint records;
• supplier attestations and audit reports.

Logs must be proportionate and protected because they may contain sensitive information. They should be time-synchronised, access-controlled, tamper-evident and retained according to a justified schedule. Excessive logging can itself create privacy and storage burdens.

Compliance is context-dependent. Organisations may need to consider data-protection law, consumer-protection rules, employment law, safety regulation, sector-specific requirements, accessibility duties, cybersecurity obligations and environmental regulation. International projects may involve conflicting legal regimes. Standards and frameworks can assist, but certification or compliance with a standard does not guarantee ethical acceptability or absence of harm.

Relevant families of guidance include information-security management standards such as ISO/IEC 27001, privacy-information management such as ISO/IEC 27701, AI management guidance such as ISO/IEC 42001, AI risk guidance such as the NIST AI Risk Management Framework, and IoT security recommendations such as ETSI EN 303 645 for consumer IoT. Organisations must verify the current edition and applicability of any standard rather than relying on outdated summaries.

8. Security as an Ethical and Governance Requirement

An insecure IoT device can expose private data, enable physical harm or become part of a wider attack. Security controls should include unique device identity, secure boot where appropriate, encrypted communication, authenticated updates, vulnerability management, least privilege, network segmentation, backup, monitoring and secure decommissioning.

Security-by-design is preferable to adding controls after deployment. Procurement should reject devices that cannot receive security updates, use hard-coded credentials or provide adequate vulnerability disclosure. The expected service life of a device should be considered: a product sold for ten years but supported for one year creates foreseeable risk.

Safety and cybersecurity interact. A malicious command, corrupted sensor or unavailable cloud service may cause physical consequences. Threat modelling should therefore include spoofing, tampering, denial of service, data poisoning, model extraction and unauthorised actuation.

9. Sustainability and Lifecycle Responsibility

IoT–AI systems consume materials, energy and infrastructure. Impacts arise from mining, semiconductor fabrication, manufacturing, packaging, transport, network operation, data centres, replacement batteries and end-of-life disposal. Sustainability must be evaluated across the lifecycle rather than by considering operational electricity alone.

Energy and carbon estimation

A simple operational energy estimate is:

\[

E = P \times t

\]

where \(E\) is energy, \(P\) is average power and \(t\) is operating time. If power is measured in watts and time in hours, energy is obtained in watt-hours; divide by 1,000 for kilowatt-hours. This relationship assumes approximately constant average power and excludes embodied energy unless it is added separately.

For example, 1,000 sensors each averaging 0.5 W and operating for 8,760 hours per year consume:

\[

E = 1{,}000 \times 0.5 \times 8{,}760 = 4{,}380{,}000\text{ Wh} = 4{,}380\text{ kWh/year}.

\]

If the relevant electricity emission factor is 0.4 kg CO₂e per kWh, operational emissions are approximately:

\[

4{,}380 \times 0.4 = 1{,}752\text{ kg CO₂e/year}.

\]

This is an illustrative estimate. Actual results depend on duty cycle, power-management states, network and cloud overhead, electricity mix and the reliability of the emission factor. It should not be presented as a complete lifecycle assessment.

AI inference and training can increase energy demand, particularly when high-frequency sensor streams are transmitted to centralised data centres. Energy reduction methods include edge processing, event-triggered sensing, model compression, quantisation, efficient architectures, renewable electricity, workload scheduling and longer device life. Energy savings must not compromise safety, accessibility or fairness.

E-waste and circular design

E-waste contains recoverable materials and hazardous substances. Short replacement cycles, sealed batteries, proprietary parts and unavailable updates encourage disposal. Circular design seeks durability, repairability, modularity, reuse, refurbishment and responsible recycling. Designers should provide replaceable batteries where safe, standard connectors, software support for the intended life, secure data wiping and material information.

A lifecycle assessment should consider functional unit, system boundary, assumptions and data quality. For example, comparing two sensors requires equal service delivered, not simply comparing the mass of the devices. A more energy-efficient device may have greater embodied impact if it is replaced frequently.

10. Responsible Innovation in Practice

Responsible innovation is anticipatory, inclusive, reflective and responsive. Teams should ask not only whether a system can be built, but whether it should be built, for whom and under what conditions.

A practical governance gate can require:

1. a clearly stated social and technical purpose;

2. alternatives analysis, including doing nothing;

3. stakeholder and rights impact assessment;

4. privacy, security, fairness and accessibility review;

5. sustainability and total-cost assessment;

6. pilot deployment with measurable stop criteria;

7. independent review for high-risk uses;

8. public or user communication in accessible language;

9. monitoring, remedy and retirement plans.

Metrics should combine technical performance with social and environmental indicators. Examples include error rates by group, complaint resolution time, percentage of devices receiving security updates, energy per useful inference, repair rate, data-retention compliance and number of unauthorised access events. Metrics can be gamed; qualitative evidence and stakeholder testimony remain important.

Applied Case: Illustrative Smart-Campus Safety System

Background

A university proposes a system combining cameras, Wi-Fi location data, emergency call points and AI-based anomaly detection to identify crowding, falls and possible violence. Alerts would be sent to campus security. The system would operate across public corridors, libraries and residence entrances.

Problem

The proposal promises faster emergency response, but it may produce false alarms, monitor students continuously and disproportionately affect groups whose movement patterns or appearance are poorly represented in training data. Residence monitoring also raises heightened privacy concerns. It is unclear who may access historical location information and how long it will be retained.

Analysis

The university first defines the purpose narrowly: detecting immediate safety incidents, not measuring attendance or student behaviour. A less intrusive alternative—manual emergency call points and limited occupancy sensors—is evaluated. A privacy impact assessment identifies location tracking and video retention as high-risk. A fairness assessment tests detection across lighting conditions, mobility aids and demographic groups, while a safety assessment examines failures caused by network loss or incorrect classification.

Stakeholder consultation includes students, staff, disability representatives, security personnel, data-protection specialists and residents. The assessment finds that continuous individual tracking is unnecessary for the stated purpose. The design is therefore changed to process most video at the edge, transmit event metadata rather than routine footage, avoid facial recognition, and use short retention periods. Access is restricted and logged. Alerts are advisory; trained personnel verify them before intervention except in narrowly defined life-threatening circumstances.

Approach and governance

The system owner approves the purpose and risk tolerance. A data steward controls retention and access. A model owner monitors false alarms and subgroup performance. Security staff receive training on automation bias and must document decisions. An independent review panel receives quarterly reports. Procurement contracts require security updates, audit access, incident notification and system shutdown if performance or privacy conditions deteriorate.

Outcome

The pilot reduces routine data collection and establishes measurable stop criteria. The university discovers that poor lighting causes more false alarms near one building and pauses deployment there until the issue is corrected. Students receive a clear explanation and a complaints route. The system is not treated as a general-purpose surveillance platform.

Learning points

The case illustrates that responsible innovation may involve reducing functionality rather than maximising data collection. Purpose limitation, data minimisation, human verification, subgroup testing, transparent governance and the ability to suspend deployment are practical controls, not abstract ideals.

Challenges and Limitations

Governance frameworks can become paperwork exercises if leaders do not provide authority, resources and consequences. Small organisations may lack specialist privacy, security or sustainability staff. Supply chains can obscure training-data provenance, model changes and component origin. Global deployments face different legal and cultural expectations.

Fairness metrics cannot resolve all value conflicts. Explainability may be limited by complex models or proprietary components. Human review can be inconsistent and may shift rather than remove responsibility. Privacy-enhancing methods can reduce utility or create new risks. Sustainability estimates often contain uncertain supplier and cloud data. Finally, standards are generally less specific than the social contexts in which systems operate.

These limitations reinforce the need for proportionality, independent challenge, continuous monitoring and engagement with affected communities. Governance should be adaptive: evidence from incidents, complaints, audits and environmental measurements must change system decisions.

Contemporary Developments and Best Practices

Current practice increasingly emphasises trustworthy AI management systems, risk-based regulation, software bills of materials, secure-by-design IoT, privacy-enhancing computation, federated learning, edge AI and lifecycle carbon accounting. These approaches are promising but not automatically responsible. Federated learning, for example, can reduce central data collection while still leaking information through updates if poorly protected. Edge processing can improve privacy and energy performance but may complicate device maintenance and auditability.

Best practices for graduates entering the field include:

• define purpose and prohibited uses before selecting technology;
• involve affected stakeholders early;
• document data, models, assumptions and limitations;
• minimise collection and retain data only as justified;
• test performance, fairness, robustness and security by relevant context;
• provide meaningful human oversight and appeal;
• maintain tamper-evident records and version control;
• monitor real-world outcomes and environmental impacts;
• procure updateable, repairable and interoperable devices;
• establish incident response, remedy and retirement procedures;
• treat compliance as a baseline, not evidence of ethical excellence.

Chapter Summary

IoT and AI create strategic opportunities but also distribute power, risk and environmental cost. Ethical systems respect autonomy, privacy, fairness, safety, accountability and sustainability. Bias may enter through data, sensors, labels, models and deployment; therefore fairness requires contextual analysis and continuing monitoring. Surveillance risks are reduced through purpose limitation, data minimisation, local processing, restricted access and meaningful consent or justified alternatives. Transparency must be tailored to individuals, operators, auditors and regulators, while human agency requires genuine competence and authority to intervene.

Governance establishes roles, decision rights, documentation, controls and oversight across the data and AI lifecycle. Auditability depends on reliable records of data, models, configurations, decisions and incidents. Standards and laws provide valuable baselines, but they do not eliminate the need for ethical judgment. Sustainability requires lifecycle analysis of energy, materials, e-waste, repairability and device longevity. Responsible innovation combines anticipation, inclusion, reflection, responsiveness and the willingness to limit or stop a system when evidence warrants it.

Key Terms

Algorithmic bias: Systematic error or disadvantage associated with data, design or deployment.

Auditability: Ability to reconstruct and assess system actions, decisions and controls.

Accountability: Obligation to explain, justify and remedy actions and impacts.

Consent: Informed and voluntary agreement to specified processing or participation.

Data minimisation: Collection and retention of no more data than necessary.

Data provenance: Information about the origin, transformation and ownership of data.

E-waste: Discarded electrical and electronic equipment and components.

Explainability: Ability to communicate understandable and relevant reasons for outputs.

Fairness: Normative requirement that people are treated justly and without unjustified discrimination.

Human agency: Capacity to understand, choose, challenge and control consequential actions.

Human-in-the-loop: Arrangement in which a person participates in the decision process.

IoT governance: Policies, roles and controls governing connected devices and data.

Lifecycle assessment: Evaluation of environmental impacts across a product or service life.

Model drift: Change in data or relationships that reduces model performance over time.

Purpose limitation: Restriction of data use to specified, legitimate purposes.

Responsible innovation: Development that anticipates impacts, includes stakeholders, reflects on values and responds to evidence.

Sustainability: Meeting present needs while protecting ecological and social capacity for the future.

Transparency: Provision of relevant information for understanding, oversight and challenge.

Review Questions

Short Answer Questions

1. Distinguish privacy from information security in an IoT system.

2. Identify four sources of algorithmic bias.

3. What makes human oversight meaningful?

4. Define auditability and name two forms of audit evidence.

5. Why is e-waste a lifecycle governance issue?

Long Answer Questions

1. Explain how data minimisation, purpose limitation and retention controls reduce surveillance risk.

2. Discuss the responsibilities of system owners, data stewards, model owners and suppliers.

3. Compare transparency requirements for individuals, operators and auditors.

4. Explain why fairness cannot be reduced to one universal statistical metric.

5. Discuss how security, safety, privacy and sustainability interact in connected devices.

Analytical / Application Questions

1. A city plans to use AI cameras to identify unusual behaviour in a transport station. Develop an ethical and governance assessment covering purpose, consent, bias, human oversight, retention and appeal.

2. A company replaces 10,000 low-power sensors every two years because batteries and software are not replaceable. Analyse the sustainability problem and propose technical and procurement changes.

3. An industrial AI system has high overall accuracy but twice the false-negative rate for one worker group. Recommend an investigation and mitigation plan, including governance, technical testing and worker participation.

Multiple Choice Questions with Answers

Question 1

Which practice best implements data minimisation in an IoT system?

A. Collecting all available sensor streams indefinitely

B. Collecting only data necessary for a defined purpose

C. Sharing raw data with every project team

D. Keeping data permanently in case it becomes useful

Correct Answer: B. Collecting only data necessary for a defined purpose

Question 2

Which is an example of measurement bias?

A. A model is updated after deployment

B. A sensor measures one demographic group less accurately

C. A dataset is encrypted at rest

D. A device uses local processing

Correct Answer: B. A sensor measures one demographic group less accurately

Question 3

Meaningful human oversight requires that the human reviewer has authority to do what?

A. Accept every automated recommendation

B. Disable all system logging

C. Reject or escalate a system output when justified

D. Remove all performance thresholds

Correct Answer: C. Reject or escalate a system output when justified

Question 4

What is the main purpose of auditability?

A. To guarantee that a model is always accurate

B. To reconstruct and assess system actions and controls

C. To eliminate the need for governance

D. To replace cybersecurity controls

Correct Answer: B. To reconstruct and assess system actions and controls

Question 5

If 1,000 sensors each use an average of 0.5 W continuously for one year of 8,760 hours, their approximate energy use is:

A. 438 kWh

B. 4,380 kWh

C. 43,800 kWh

D. 438,000 kWh

Correct Answer: B. 4,380 kWh

Question 6

Which control most directly supports purpose limitation?

A. Using collected data for unrelated future purposes without review

B. Defining approved uses and restricting secondary use

C. Increasing sensor sampling frequency

D. Removing all access logs

Correct Answer: B. Defining approved uses and restricting secondary use

Question 7

Why is compliance with a standard not sufficient to prove ethical acceptability?

A. Standards never contain technical requirements

B. Ethical impacts may exceed the scope of a standard

C. Standards prohibit all independent audits

D. Compliance prevents stakeholder consultation

Correct Answer: B. Ethical impacts may exceed the scope of a standard

Question 8

Which design choice generally supports circularity?

A. Sealed batteries and proprietary connectors

B. Short software-support periods

C. Modular, repairable hardware with secure data wiping

D. Frequent replacement of complete devices

Correct Answer: C. Modular, repairable hardware with secure data wiping

Question 9

A feedback loop occurs when:

A. A model output influences behaviour and generates reinforcing future data

B. A device encrypts its network traffic

C. A sensor is physically calibrated

D. An auditor reviews a model card

Correct Answer: A. A model output influences behaviour and generates reinforcing future data

Question 10

In the illustrative smart-campus case, why was routine facial recognition rejected?

A. It was required for all emergency systems

B. It was unnecessary for the narrowly defined safety purpose and increased privacy risk

C. It guaranteed perfect detection

D. It reduced the need for governance

Correct Answer: B. It was unnecessary for the narrowly defined safety purpose and increased privacy risk