Ethics, Governance and Future Directions

Introduction

Text and social media analytics can reveal public concerns, identify emerging risks, evaluate campaigns, support customer service, and inform policy. The same capabilities can also enable intrusive surveillance, discriminatory profiling, political manipulation, copyright infringement, and opaque automated decisions. Ethical practice is therefore not an optional supplement to analytical skill. It is a condition of professional competence.

The central governance problem is a tension between analytical value and individual and social rights. Textual data may appear public, yet public availability does not remove contextual expectations of privacy. A post may be technically accessible but intended for a limited community, a particular audience, or a specific conversational purpose. Ethical analytics must therefore ask not only “Can we collect this data?” but also “Should we collect it, for what purpose, with what safeguards, and with what consequences?”

This chapter develops a practical framework for responsible text and social media analytics. It examines consent, privacy, ownership, surveillance, fairness, transparency, content moderation, platform policies, regulation, and organizational governance. It then considers multimodal AI, agentic systems, real-time analytics, and likely future directions.

Learning Objectives

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

1. Distinguish legal compliance, ethical responsibility, and organizational governance.
2. Evaluate consent, privacy, ownership, and surveillance risks in text and social media projects.
3. Apply fairness, transparency, accountability, and data-minimization principles to analytical systems.
4. Critically assess content moderation and platform-policy decisions.
5. Design a governance process for an analytics project, including impact assessment and audit controls.
6. Evaluate emerging risks associated with multimodal AI, agentic systems, and real-time analytics.

1. Ethics, Law and Governance: Related but Different

Ethics concerns what ought to be done. It includes duties to respect autonomy, avoid harm, promote fairness, preserve dignity, and distribute benefits and burdens responsibly. Ethical obligations may exceed legal minimums.

Law consists of enforceable rules established by legislatures, regulators, courts, and recognized authorities. Relevant legal regimes vary by jurisdiction and may include data-protection law, consumer-protection law, employment law, intellectual-property law, sector-specific confidentiality rules, and rules concerning automated decision-making.

Governance is the institutional system through which decisions are authorized, monitored, documented, challenged, and corrected. Governance includes policies, roles, approval procedures, technical controls, audits, incident response, and accountability mechanisms.

A lawful system can still be unethical. For example, a company may collect publicly visible posts under a broad contractual permission but use them to infer sensitive health or political characteristics without reasonable user expectations. Conversely, an ethically desirable practice may require controls more demanding than the law explicitly specifies.

A useful governance principle is:

The greater the potential impact on people, the stronger the justification, oversight, transparency, and contestability required.

A risk-based view

Risk is not determined solely by the volume of data. A small dataset can create severe risk if it concerns vulnerable people or sensitive attributes. A practical assessment considers:

• Sensitivity: health, sexuality, religion, political belief, location, or financial information.
• Identifiability: whether individuals can be directly or indirectly re-identified.
• Scale: number of people, frequency, and geographic reach.
• Inference power: whether apparently harmless data can reveal sensitive traits.
• Impact: consequences for employment, credit, education, safety, reputation, or liberty.
• Vulnerability: age, social position, language, disability, or exposure to coercion.
• Reversibility: whether harm can be corrected once an output is published or acted upon.

2. Consent, Purpose and Context

Consent is meaningful only when it is informed, specific, freely given, and capable of withdrawal where applicable. A long privacy notice or a general platform registration agreement does not automatically provide meaningful consent for every downstream analytical use.

Researchers and organizations should distinguish at least four situations:

1. Directly solicited data: participants knowingly provide responses for a defined study or service.
2. Platform-mediated data: users post under a platform’s terms, but the analyst is a separate party with a different purpose.
3. Observed data: content is collected from public or semi-public spaces without direct interaction.
4. Derived data: labels, embeddings, sentiment scores, risk scores, or inferred attributes are generated from original content.

Derived data deserve special attention. An organization may avoid storing a person’s explicit health statement yet produce a health-risk score from language patterns. The absence of the original sensitive field does not eliminate sensitivity in the inference.

Purpose limitation

Purpose limitation requires that data be collected and used for specified, legitimate purposes. “Collect everything now and decide later” is difficult to defend because it increases exposure, weakens accountability, and encourages function creep. A project proposal should state:

• the decision or research question;
• the minimum data required;
• the intended users and outputs;
• retention and deletion rules;
• prohibited secondary uses;
• procedures for new uses or scope changes.

Consent is not the only possible legal basis in all jurisdictions; other bases may include contract, legal obligation, public interest, or legitimate interest. However, selecting a legal basis does not remove duties of fairness, transparency, security, proportionality, and respect for rights. Legal advice is required for jurisdiction-specific decisions.

3. Privacy, Anonymity and Re-identification

Privacy is broader than secrecy. It includes control over personal information, freedom from unjustified observation, contextual integrity, and protection against harmful use. Public data can remain personal data when it relates to an identifiable person.

Anonymization aims to make identification sufficiently unlikely and irreversible in the relevant context. Pseudonymization replaces identifiers with codes but retains a possibility of re-linking; it is therefore a security and risk-reduction measure, not equivalent to anonymity.

Re-identification can occur through:

• combining several datasets;
• distinctive phrases searchable on the web;
• timestamps and locations;
• social-network structure;
• writing style or authorship attribution;
• rare demographic combinations;
• model outputs that memorize training examples.

A simple privacy-risk expression sometimes used in governance is expected loss:

Here, p is the estimated probability of a harmful event and I is its impact. This is a decision aid, not an objective measurement of ethics. Probability estimates may be uncertain, and severe rights violations should not be accepted merely because their average expected loss appears low. The expression is useful for comparing controls under stated assumptions; it does not replace qualitative human-rights analysis.

Privacy-preserving practices

Appropriate controls may include data minimization, aggregation, removal of direct identifiers, access restrictions, encryption, secure environments, retention limits, query logging, differential privacy, and contractual restrictions on re-use. Differential privacy provides a formal way to limit how much an individual’s inclusion changes a released statistic. Its privacy parameter must be selected carefully: stronger privacy generally reduces analytical precision, and poor implementation can still reveal information through repeated queries or auxiliary data.

4. Ownership, Intellectual Property and Data Rights

Ownership of a social-media post is not a single, universal question. Relevant issues include:

• copyright in the wording, image, video, or audio;
• database rights where recognized;
• platform licenses granted by users;
• contractual restrictions on scraping and API use;
• publicity, personality, and moral rights;
• confidentiality and trade secrets;
• rights in annotations, labels, and model outputs.

A platform’s permission to host content does not necessarily grant an analyst unrestricted rights to copy, redistribute, train a commercial model, or publish verbatim quotations. Terms of service may also prohibit automated collection or impose rate, security, and deletion requirements. Copyright exceptions and text-and-data-mining rules differ across jurisdictions and fact patterns.

A responsible project keeps a data provenance record identifying source, collection method, permission or legal basis, transformations, licenses, and downstream uses. Publishing paraphrased or aggregated findings may reduce exposure, but paraphrasing does not guarantee that an individual cannot be identified or that intellectual-property concerns disappear.

5. Surveillance and Power

Analytics becomes surveillance when observation is systematic, persistent, and used to influence, classify, or control people. Surveillance concerns arise in workplace monitoring, student monitoring, border and security contexts, political communication, advertising, and location-aware services.

The ethical issue is not merely whether monitoring is accurate. It also concerns power asymmetry, reasonable expectations, chilling effects, proportionality, and the possibility of automated escalation. Employees or citizens may technically agree to monitoring while lacking a realistic ability to refuse.

A proportionality assessment should ask:

1. Is the objective legitimate and sufficiently important?
2. Is text or social-media analytics necessary to achieve it?
3. Is there a less intrusive method?
4. Are the benefits proportionate to the intrusion and risks?
5. Are access, retention, and appeal controls adequate?

Continuous monitoring can alter behavior even when no formal sanction occurs. This chilling effect is a social cost that standard model-performance metrics do not capture.

6. Fairness, Bias and Discrimination

Bias can enter through historical behavior, sampling, labels, language coverage, annotation practices, platform ranking, missing data, proxy variables, and deployment context. A sentiment classifier trained mostly on one dialect may systematically misclassify another. A toxicity model may treat identity terms as abusive because they co-occur with harassment in training data. A crisis detector may overlook communities whose language is underrepresented.

Fairness is multidimensional. Possible concerns include:

• Representational fairness: groups are not erased, stereotyped, or misrepresented.
• Allocative fairness: benefits and burdens are not unjustly distributed.
• Procedural fairness: people receive consistent treatment and meaningful recourse.
• Performance fairness: error patterns are not unacceptably unequal across groups.
• Intersectional fairness: combinations of attributes are considered rather than only broad averages.

No single fairness metric resolves every conflict. For a binary classifier, common group metrics include:

For group g, TP_g, FN_g, FP_g, and TN_g denote true positives, false negatives, false positives, and true negatives. TPR_g is the true-positive rate and FPR_g is the false-positive rate. These rates apply to a defined binary task with reliable labels and a specified decision threshold. Equalizing them may be impossible simultaneously with other criteria when groups have different base rates. Metrics should therefore support—not replace—contextual judgment, stakeholder consultation, and examination of actual harms.

Fairness assessment should occur before deployment, during monitoring, and after material changes. Teams should disaggregate performance by relevant language, region, disability, age, gender, and other legally and ethically appropriate categories, while avoiding unnecessary collection of sensitive data.

7. Transparency, Explainability and Accountability

Transparency means making relevant information available. It may include the purpose of processing, data sources, model capabilities and limitations, responsible personnel, retention, appeal routes, and significant changes. Transparency does not require disclosing security-sensitive details or proprietary code in every case.

Explainability concerns how a system’s output can be understood. A global explanation describes general model behavior; a local explanation addresses one prediction. Neither automatically establishes that a decision is lawful or fair. A plausible explanation can be misleading if the model relies on unstable correlations.

A governance documentation package should include:

Accountability requires an identifiable owner with authority and resources. “The algorithm decided” is not an acceptable allocation of responsibility.

8. Content Moderation and Platform Governance

Content moderation combines policy interpretation, detection, ranking, human review, enforcement, and appeal. It is not simply a classification problem. A post may be offensive but permitted, harmful but ambiguous, or dangerous only in a particular context. Satire, quotation, counterspeech, journalism, and reclaimed language create difficult cases.

Moderation systems commonly use a layered process:

1. define a policy and protected exceptions;
2. detect potentially violating content;
3. apply risk-sensitive triage;
4. conduct human or specialized review for uncertain or high-impact cases;
5. impose proportionate action;
6. notify the user where appropriate;
7. provide appeal and correction;
8. audit outcomes across languages and groups.

Automated removal may be appropriate for high-confidence, high-severity material, but borderline cases require context. Over-removal can suppress lawful speech, political participation, or minority expression. Under-removal can expose users to harassment, incitement, exploitation, or coordinated manipulation.

Platform governance should publish intelligible rules, enforcement statistics, error information, researcher-access procedures, and appeal outcomes where feasible. It should also address recommender systems, since amplification can create harm even when the underlying content is not prohibited.

9. Regulation and Organizational Governance

Regulation is increasingly risk-based. Contemporary regimes and standards commonly emphasize transparency, data protection, security, accountability, risk assessment, human oversight, and restrictions on high-risk or harmful applications. Requirements differ by location and sector, so organizations must map the jurisdictions in which data subjects, users, infrastructure, and decision consequences occur.

A governance operating model can be organized into three lines:

• First line—product and analytics teams: build safeguards, document choices, monitor performance, and own operational risk.
• Second line—privacy, legal, security, risk, and ethics functions: advise, challenge, approve, and set controls.
• Third line—internal audit or independent assurance: test whether controls work in practice.

A project impact assessment should cover data flows, affected populations, legal basis, foreseeable harms, model limitations, human oversight, security, retention, vendor dependencies, and exit or deletion plans. High-impact uses should receive approval before deployment, not after an incident.

10. A Practical Responsible Analytics Lifecycle

Responsible analytics is a lifecycle rather than a one-time checklist.

Stage 1: Define the purpose

Specify the decision, beneficiaries, affected people, alternatives, and unacceptable uses. Reject projects whose purpose is vague or whose expected benefit cannot justify the intrusion.

Stage 2: Map data and stakeholders

Create a data-flow map showing collection, storage, transformation, model access, outputs, vendors, and deletion. Consult users, affected communities, domain experts, legal specialists, and security professionals.

Stage 3: Assess risk and design controls

Evaluate privacy, fairness, security, intellectual property, safety, and societal risks. Select minimization, access, testing, human-review, and appeal controls.

Stage 4: Build and validate

Use representative evaluation data, document labeling decisions, test subgroup performance, conduct adversarial and privacy testing, and preserve reproducibility. Do not use synthetic data as a substitute for understanding real-world coverage; synthetic data can reproduce the biases of its generator.

Stage 5: Deploy cautiously

Begin with a limited pilot, define stop conditions, separate recommendations from automatic decisions, and ensure trained human oversight. Monitor drift in language, events, platform behavior, and user populations.

Stage 6: Monitor, contest and improve

Track complaints, false positives, disparate impacts, security events, model changes, and unintended uses. Provide accessible correction and appeal. Retire systems when risks exceed benefits or when the purpose ends.

11. Illustrative Case: CivicPulse Election Conversation Monitor

Background

CivicPulse, a fictional public-interest organization, proposes a dashboard that analyzes public posts during an election to identify emerging concerns and misinformation themes. It plans to collect posts in several languages, classify topics and toxicity, and provide daily summaries to journalists and election administrators.

Problem

The initial proposal would store verbatim posts indefinitely, display links to individual accounts, infer political affiliation, and automatically label content as misinformation. Smaller-language communities are poorly represented in the training data. The team argues that all posts are public and that rapid action is essential.

Analysis

The project has legitimate public-interest value but creates several risks. Public accessibility does not remove privacy expectations. Political affiliation is sensitive in many contexts, and inferred affiliation can expose people to retaliation. Permanent storage increases re-identification risk. A binary misinformation label may confuse disputed claims, satire, opinion, and evolving evidence. Language imbalance may produce unequal false positives. Direct account links could turn aggregate monitoring into individual surveillance.

Approach and solution

The governance board redesigns the project as follows:

• collect only posts necessary for defined election-risk indicators;
• minimize or hash account identifiers and avoid publishing handles;
• retain verbatim content only for a short, justified validation period;
• publish aggregate trends and representative paraphrases rather than searchable quotations;
• separate topic detection from factual verification;
• require expert review before a claim is labeled false;
• report uncertainty and language-specific performance;
• use human review for high-impact escalation;
• create a correction channel for users and civil-society organizations;
• document platform permissions and prohibit use for voter profiling;
• conduct weekly bias, privacy, and security reviews during the election period;
• establish deletion and shutdown triggers.

Outcome

The redesigned system delivers slower but more defensible public-interest reporting. It identifies coordinated narratives at aggregate level while reducing individual exposure. Some minority-language outputs are withheld until evaluation quality improves. The project does not claim to determine truth automatically; it supplies evidence for qualified human investigation.

Learning points

The case demonstrates that responsible design often changes the objective, data resolution, retention period, and publication format—not merely the model. It also shows why speed does not eliminate the need for governance; rapid decisions make uncertainty and appeal mechanisms more important.

12. Future Directions

Multimodal AI

Text analytics increasingly operates with images, video, audio, metadata, and interaction patterns. Multimodal models can interpret memes, spoken language, visual context, and coordinated campaigns more effectively than text-only systems. They also increase risk: biometric identification, voice imitation, sensitive visual inference, copyright disputes, and cross-modal re-identification become possible.

Governance should treat each modality as a distinct source of sensitivity. A harmless caption combined with a face image and location may become highly identifying. Evaluation should test cultural context, accessibility, dialect, visual ambiguity, and adversarial content. Consent and deletion must cover derived representations such as embeddings, not only original files.

Agentic systems

An agentic system can plan tasks, call tools, retrieve data, send messages, and take actions with limited step-by-step human direction. In social-media analytics, an agent might detect a trend, query multiple sources, draft a response, and trigger a campaign.

The risk is not only inaccurate output but compound failure: a small error in retrieval can lead to a mistaken classification, an inappropriate recommendation, and an irreversible public action. Safe deployment requires least-privilege tool access, action limits, approval gates, provenance, sandboxing, continuous logs, and an emergency stop. Agents should not autonomously target individuals, publish accusations, or make high-impact decisions without accountable human authorization.

Real-time analytics

Streaming systems can detect crises and coordinated abuse quickly, but speed creates pressure to act on incomplete evidence. Real-time governance should define latency tiers: automatic low-risk alerts, analyst review for moderate-risk events, and senior authorization for high-impact interventions. Systems need backfilling, correction, replay, and post-event audit capabilities. A model optimized for immediate detection may generate more false positives; threshold choices should reflect the relative costs of missed and erroneous interventions.

Synthetic media and provenance

Generative systems make it easier to create persuasive text, images, audio, and video at scale. Detection alone is unlikely to solve the problem because detectors can fail under distribution shift and adversarial adaptation. Complementary measures include provenance metadata, source authentication, media literacy, independent verification, rate limits, identity protection, and clear labeling. Provenance signals should be treated as evidence rather than infallible truth.

Privacy-enhancing computation and federated approaches

Organizations may increasingly use federated learning, secure computation, access-controlled clean rooms, and privacy-preserving statistics to reduce centralized data exposure. These methods can improve governance but introduce complexity, implementation risk, and potential loss of interpretability. They are safeguards, not ethical exemptions.

Strategic implications for managers

Managers should treat governance as a capability that creates durable analytical value. Procurement should require vendor documentation, incident notification, audit access, data deletion, subcontractor disclosure, and restrictions on model training. Boards should ask who is affected, who can challenge a decision, what happens when the model is wrong, and whether the organization can stop the system safely.

Challenges and Limitations

Responsible analytics faces persistent difficulties. Laws differ across jurisdictions and change rapidly. Consent can be impractical in large-scale public-data research, while alternative legal bases may be contested. Fairness metrics can conflict, and sensitive-group evaluation may itself create privacy risks. Explainability methods may be unstable. Platform APIs and policies can change without notice. Vendors may provide incomplete information. Human review can introduce inconsistency, fatigue, and its own biases. Finally, governance cannot eliminate all uncertainty or misuse; it can only make risks more visible, constrained, and accountable.

The appropriate response is not to abandon analytics but to align ambition with evidence, necessity, proportionality, and institutional responsibility.

Best-Practice Checklist

Before approval, an analytics team should be able to answer “yes” to the following:

• Is the purpose specific, legitimate, and necessary?
• Are data collection and retention minimized?
• Are consent, legal basis, platform terms, and intellectual-property issues documented?
• Have affected communities and vulnerable groups been considered?
• Have subgroup performance, privacy, and security risks been tested?
• Are outputs uncertain where evidence is uncertain?
• Is a responsible human owner identified?
• Can people correct, appeal, or obtain meaningful information?
• Are high-impact actions gated by human review?
• Are monitoring, incident response, deletion, and shutdown procedures operational?

Chapter Summary

Text and social media analytics creates substantial public and commercial value but can also intensify privacy invasion, surveillance, discrimination, manipulation, and intellectual-property conflict. Ethical practice extends beyond legal compliance and requires purpose limitation, proportionality, data minimization, security, fairness, transparency, accountability, and contestability. Public data are not automatically free of privacy obligations, and derived inferences may be more sensitive than source content.

Content moderation is a socio-technical governance activity involving policy, context, enforcement, appeals, and recommender effects. A responsible analytics lifecycle begins with a defined purpose, maps data and stakeholders, assesses risks, validates subgroup performance, deploys cautiously, and continuously monitors and corrects outcomes. Future systems involving multimodal AI, agents, real-time streams, and synthetic media require stronger controls because they combine scale, speed, autonomy, and increasingly persuasive outputs. The durable professional principle is to design systems that are useful, proportionate, explainable, contestable, and capable of being stopped.

Key Terms

Accountability: Responsibility for decisions, controls, outcomes, and remediation.

Anonymization: Processing intended to make identification sufficiently unlikely and irreversible in context.

Bias: Systematic distortion that can produce unequal or inaccurate outcomes.

Consent: Informed and voluntary agreement to a specified processing activity.

Content moderation: Detection, review, enforcement, and appeal concerning platform content.

Contextual integrity: Privacy expectation based on the norms of a particular social context.

Data minimization: Limiting collection and retention to what is necessary.

Differential privacy: A formal privacy approach limiting the effect of one individual’s data on a released result.

Explainability: The extent to which system behavior or an output can be understood.

Fairness: Justifiable treatment and distribution of errors, benefits, and burdens.

Function creep: Expansion of data use beyond the original purpose.

Governance: Structures and processes for authorization, oversight, correction, and accountability.

Inference: Derivation of attributes or conclusions from observed data.

Multimodal AI: AI that processes or generates more than one modality, such as text, image, audio, or video.

Pseudonymization: Replacement of identifiers with codes while retaining a re-linking possibility.

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

Re-identification: Linking supposedly de-identified data to a person or entity.

Surveillance: Systematic observation used to understand, classify, influence, or control.

Transparency: Provision of relevant information about purposes, processes, limitations, and responsibility.

Agentic system: A system capable of planning and taking tool-mediated actions with limited step-by-step direction.

Review Questions

Short Answer Questions

1. Why is public availability of a social-media post not sufficient to resolve privacy concerns?
2. Distinguish anonymization from pseudonymization.
3. What is purpose limitation, and why does it matter?
4. Identify two ways bias can enter a text-classification system.
5. Why are appeals important in content moderation?

Long Answer Questions

1. Explain the differences and relationships among ethics, law, and governance in text and social media analytics.
2. Discuss how proportionality can guide decisions about workplace or civic surveillance.
3. Compare representational, allocative, procedural, and performance fairness.
4. Describe a responsible analytics lifecycle from project definition through retirement.
5. Evaluate the governance challenges created by multimodal and agentic AI.

Analytical / Application Questions

1. A retailer wants to infer customers’ health conditions from public posts to personalize advertising. Assess the proposal using consent, purpose limitation, privacy, fairness, and proportionality principles. Recommend a defensible alternative or explain why the project should not proceed.
2. A platform’s toxicity model removes posts containing identity terms at a higher rate for one language community. Design an investigation and remediation plan, including data, metrics, human review, appeals, and monitoring.
3. An election-monitoring agent can collect posts, classify claims, and publish alerts without approval. Design an authorization and control architecture that preserves timely detection while preventing unjustified public accusations.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes law from ethics in analytics governance?

A. Law concerns only technical accuracy, whereas ethics concerns data storage.

B. Law consists of enforceable rules, while ethics also concerns what ought to be done beyond legal minimums.

C. Ethics is enforceable by courts, while law is voluntary.

D. There is no meaningful distinction between law and ethics.

Correct Answer: B. Law consists of enforceable rules, while ethics also concerns what ought to be done beyond legal minimums.

Question 2

Which practice most directly implements purpose limitation?

A. Collecting all available posts for possible future use

B. Retaining all raw data indefinitely

C. Defining the decision question and prohibiting unrelated secondary uses

D. Publishing individual usernames with every result

Correct Answer: C. Defining the decision question and prohibiting unrelated secondary uses

Question 3

What is the main difference between pseudonymization and anonymization?

A. Pseudonymization always makes data public.

B. Anonymization requires retaining a secret key.

C. Pseudonymized data may still be re-linked to a person.

D. They are legally and technically identical.

Correct Answer: C. Pseudonymized data may still be re-linked to a person.

Question 4

In the formula TPR_g = TP_g ÷ (TP_g + FN_g), what does TPR_g represent?

A. False-positive rate

B. True-positive rate

C. Overall accuracy

D. Prevalence

Correct Answer: B. True-positive rate

Question 5

Why can a public post still create privacy risk?

A. Public posts are always encrypted.

B. Public visibility eliminates all expectations of privacy.

C. Posts may be combined with other data to identify or profile individuals.

D. Privacy applies only to offline information.

Correct Answer: C. Posts may be combined with other data to identify or profile individuals.

Question 6

Which is the strongest example of accountability?

A. Allowing an algorithm to make unreviewable decisions

B. Assigning a responsible owner and maintaining audit and appeal processes

C. Removing all documentation to protect trade secrets

D. Using a larger dataset without monitoring

Correct Answer: B. Assigning a responsible owner and maintaining audit and appeal processes

Question 7

Why are borderline content-moderation cases often unsuitable for automatic removal?

A. They never involve harm.

B. Context, satire, quotation, and counterspeech can change interpretation.

C. Automated systems cannot process any text.

D. Users cannot appeal automated decisions.

Correct Answer: B. Context, satire, quotation, and counterspeech can change interpretation.

Question 8

Which control is especially important for an agentic system that can publish content?

A. Unlimited tool access

B. Removal of logs

C. Approval gates and least-privilege permissions

D. Permanent autonomous operation

Correct Answer: C. Approval gates and least-privilege permissions

Question 9

What is a key risk of real-time analytics?

A. It always produces perfectly complete evidence.

B. It can encourage action on incomplete or rapidly changing information.

C. It eliminates false positives.

D. It cannot be used for alerts.

Correct Answer: B. It can encourage action on incomplete or rapidly changing information.

Question 10

Which statement best describes fairness metrics?

A. One metric resolves every ethical question.

B. Metrics are useful evidence but must be interpreted in context and may conflict.

C. Metrics make stakeholder consultation unnecessary.

D. Fairness requires ignoring subgroup performance.

Correct Answer: B. Metrics are useful evidence but must be interpreted in context and may conflict.