Ethics, Privacy, Security and Data Governance

Introduction and Context

Data analytics creates value by converting observations into decisions, but analytical capability also creates risks. A model may discriminate, reveal sensitive attributes, enable surveillance, expose confidential information, or produce decisions that cannot be explained or challenged. Consequently, responsible analytics is not limited to compliance after a model has been built. It requires an integrated system of ethical reasoning, privacy protection, information security, governance, model validation, documentation, and human accountability.

This chapter treats analytics as a socio-technical activity. The quality of an analytical decision depends not only on predictive accuracy but also on whether the data were legitimately obtained, whether affected individuals were treated fairly, whether security controls are proportionate to risk, whether decisions are explainable, and whether the organisation can demonstrate what happened and why.

Learning Objectives

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

1. Distinguish ethics, privacy, security, compliance, and data governance.

2. Apply principles of consent, purpose limitation, minimisation, retention, and individual rights.

3. Compare major legal and governance frameworks relevant to analytics.

4. Assess fairness, explainability, robustness, auditability, and model risk.

5. Design governance roles, controls, approval processes, and incident responses.

6. Evaluate an analytical system through a risk-based responsible-AI lifecycle.

1. Foundations of Responsible Analytics

Ethics concerns what an organisation ought to do, including duties to respect autonomy, prevent harm, promote fairness, and preserve human dignity. Privacy concerns appropriate control over information about identifiable or identifiable persons. Security protects information and systems against unauthorised access, alteration, disclosure, disruption, or destruction. Compliance concerns conformity with applicable law, regulation, contracts, and internal policy. Data governance is the allocation of decision rights, responsibilities, standards, and controls throughout the data lifecycle.

These concepts overlap but are not interchangeable. An activity can be legal but ethically questionable; secure data can still have been collected without meaningful consent; and a compliant model can still disadvantage a protected group. Governance provides the organisational mechanism for aligning these dimensions.

A useful risk formulation is:

Risk = likelihood of harm × impact of harm

This is a conceptual risk-ranking relationship rather than a precise probability equation. Likelihood and impact should be assessed using defined scales, evidence, and documented assumptions. High-impact uses, such as credit, employment, healthcare, education, policing, and insurance, generally require stronger controls than low-impact recommendation systems.

Ethical principles

• Autonomy and informed choice: people should understand significant uses of their data and retain meaningful choices where feasible.
• Beneficence and non-maleficence: analytics should generate legitimate benefit while preventing foreseeable harm.
• Justice and fairness: benefits and burdens should not be distributed arbitrarily or systematically disadvantage vulnerable groups.
• Accountability: an identifiable person or body must be answerable for outcomes.
• Transparency: relevant information about data, methods, limitations, and decisions should be communicated.
• Proportionality: the intrusiveness and control burden should be proportionate to the purpose and risk.

2. Privacy Across the Data Lifecycle

Privacy protection begins before collection and continues through deletion. A lifecycle approach includes collection, preparation, analysis, sharing, deployment, retention, and disposal.

Lawful and fair collection

An organisation should identify a lawful basis or authorisation, define the purpose, provide an intelligible notice, and collect only data necessary for that purpose. Consent is not automatically valid merely because a person clicked an acceptance box. Meaningful consent should be informed, specific, freely given where applicable, and withdrawable. Where there is a power imbalance, such as employment or public services, consent may not be genuinely voluntary; another legal basis or a different service design may be required.

Purpose limitation prevents data collected for one purpose from being silently repurposed for an incompatible one. Data minimisation means collecting and retaining what is necessary, not everything that might eventually be useful. Minimisation reduces both privacy exposure and attack surface.

Identifiability and de-identification

Direct identifiers include names, account numbers, and email addresses. Quasi-identifiers, such as date of birth, postcode, occupation, and rare events, can identify a person when combined. Removing names therefore does not necessarily anonymise a dataset.

Pseudonymisation replaces identifiers with codes but preserves a linkage mechanism; the data remain potentially personal. Anonymisation aims to make re-identification reasonably unlikely using available means, although no technique guarantees permanent anonymity in all future contexts.

A k-anonymity example illustrates the idea. A dataset is 3-anonymous with respect to selected quasi-identifiers if every combination of those attributes appears at least three times. This can reduce simple linkage risk, but it does not prevent attribute disclosure when all three records share the same sensitive value. Differential privacy provides a stronger formal approach by adding calibrated randomness. For a mechanism M and neighbouring datasets D and D' differing in one person, ε-differential privacy requires, for every output set S:

Pr[M(D) ∈ S] ≤ e^ε Pr[M(D') ∈ S]

Here ε ≥ 0 is the privacy-loss parameter; smaller ε generally provides stronger privacy but may reduce analytical utility. The guarantee depends on the mechanism, the neighbouring-dataset definition, composition across repeated queries, and correct privacy-budget accounting. It is not a licence to publish arbitrary outputs without auxiliary safeguards.

Individual rights and retention

Depending on the jurisdiction and context, people may have rights of access, correction, deletion, restriction, objection, portability, or review of automated decisions. Organisations need operational processes, not merely statements in a privacy notice. Retention schedules should specify purpose, owner, period, archival conditions, and secure deletion method. Keeping data indefinitely because storage is inexpensive is not a defensible governance policy.

3. Security for Analytical Systems

Analytics environments contain raw data, derived features, credentials, model artefacts, prompts, logs, and outputs. Controls should address confidentiality, integrity, and availability—the CIA triad—while recognising that model integrity and output misuse are additional concerns.

Core controls include:

• identity verification and least-privilege access;
• role-based or attribute-based authorisation;
• encryption in transit and at rest, with managed key rotation;
• tokenisation or masking for sensitive fields;
• network segmentation and secure development practices;
• vulnerability management, patching, backup, and recovery testing;
• immutable or access-controlled audit logs;
• data-loss prevention and monitoring for unusual extraction;
• incident response, notification, and post-incident learning.

The principle of zero trust treats every request as requiring verification rather than assuming that an internal network is safe. In analytical pipelines, security must cover notebooks, cloud storage, orchestration services, application programming interfaces, third-party data, and model-serving endpoints.

AI systems introduce distinctive threats. Training-data poisoning can corrupt the learned behaviour; prompt injection can manipulate a connected language model; membership inference can reveal whether an individual was in training data; model inversion can infer sensitive characteristics; and adversarial examples can induce incorrect classifications. Controls include provenance checks, dataset validation, red-team testing, output filtering, rate limits, access isolation, monitoring, and human escalation.

4. Regulatory and Governance Frameworks

Legal obligations are jurisdiction- and sector-dependent. A responsible analytics programme begins with a data inventory, a processing map, jurisdiction analysis, and legal review rather than assuming that one global policy is sufficient.

GDPR and the European data-protection model

The EU General Data Protection Regulation (GDPR) applies broadly to personal-data processing within its territorial scope and, in some cases, to organisations outside the EU offering services to or monitoring people in the EU. Important concepts include lawfulness, fairness and transparency; purpose limitation; data minimisation; accuracy; storage limitation; integrity and confidentiality; and accountability. It distinguishes controllers from processors, regulates international transfers, recognises data-subject rights, and requires data-protection impact assessments for processing likely to create high risk. Certain solely automated decisions with legal or similarly significant effects receive special protection, subject to the regulation’s conditions and exceptions.

The GDPR is principles-based and accountability-oriented. It does not make accuracy alone sufficient: necessity, proportionality, rights, lawful basis, and demonstrable controls also matter. A data-protection officer may be required in specified circumstances.

CCPA/CPRA and the United States privacy landscape

California’s CCPA, as amended by the CPRA, gives qualifying California residents rights including access, deletion, correction, opting out of sale or sharing, and limiting certain uses of sensitive personal information. It distinguishes business, service-provider, and contractor relationships and imposes notice and contractual requirements. Unlike the GDPR, the US framework is fragmented: sectoral laws such as HIPAA for protected health information, GLBA for certain financial institutions, and FERPA for specified education records operate alongside state privacy laws and consumer-protection principles. Organisations must therefore determine which sectoral and state rules apply rather than describing “US privacy law” as one system.

HIPAA and health analytics

HIPAA regulates covered entities and business associates in relation to protected health information. Its Privacy Rule governs permitted uses and disclosures, while the Security Rule requires administrative, physical, and technical safeguards. The minimum-necessary principle limits access and disclosure to what is needed for the purpose. De-identification may be established through expert determination or specified removal of identifiers, but organisations must still manage re-identification and contractual risks.

EU AI Act and risk-based AI regulation

The EU AI Act adopts a risk-based structure that prohibits certain practices, imposes strict requirements on high-risk systems, and creates transparency obligations for specified systems. High-risk providers and deployers may face requirements concerning risk management, data governance, technical documentation, record keeping, accuracy, robustness, cybersecurity, human oversight, and post-market monitoring. Classification depends on use and context, not simply on whether a system uses machine learning.

Standards and guidance

The NIST AI Risk Management Framework is a voluntary framework organised around governing, mapping, measuring, and managing AI risks. ISO/IEC 27001 addresses information-security management systems; ISO/IEC 27701 extends privacy-information management; and ISO/IEC 42001 specifies an AI management system. These standards are not automatically law, but they provide structured control language, evidence expectations, and continual-improvement mechanisms. Certification or alignment does not guarantee ethical outcomes; controls must be appropriate to the actual use.

Comparative implications

The GDPR emphasises rights, lawful processing, and accountability; CCPA/CPRA emphasises consumer rights and disclosures within a broader state-based system; HIPAA is sector-specific; the EU AI Act focuses on AI-use risk categories; NIST and ISO provide management and control frameworks. A multinational organisation should maintain a regulatory obligations register mapping each use case to jurisdiction, data category, affected persons, legal basis, retention, rights process, breach duties, transfer restrictions, and accountable owner.

5. Fairness and Discrimination

Bias can enter through historical inequity, under-representation, measurement error, label choices, sampling, proxy variables, deployment context, or unequal access to human appeal. Removing a protected attribute does not remove bias because correlated variables may act as proxies.

Let TP, TN, FP, and FN denote true positives, true negatives, false positives, and false negatives. Common group metrics include:

• False-positive rate (FPR): FP/(FP + TN).
• False-negative rate (FNR): FN/(FN + TP).
• Positive predictive value: TP/(TP + FP).
• Demographic parity: P(Ŷ = 1 | A = a) is equal across groups.
• Equal opportunity: true-positive rates are equal across groups.
• Equalised odds: both true-positive and false-positive rates are equal across groups.

These metrics can conflict, especially when groups have different base rates or score distributions. Therefore, fairness is not achieved by selecting a metric mechanically. Teams must identify the relevant harm, justify the metric, examine intersectional groups, assess calibration and error costs, and provide remediation and appeal mechanisms. Fairness metrics are evidence for a governance decision, not substitutes for that decision.

6. Explainability, Transparency and Contestability

Interpretability concerns how understandable a model is to humans. Explainability concerns methods used to describe a model or a particular output. Global explanations describe overall behaviour; local explanations describe one prediction. Feature importance, partial-dependence plots, counterfactual explanations, surrogate models, and inherently interpretable models each have limitations.

A local explanation is not proof of causation. Post-hoc explanations may be unstable, incomplete, or misleading when features are correlated. High-stakes systems should combine technical explanation with process transparency: the purpose, data categories, decision authority, material factors, limitations, human reviewer, and challenge route should be documented. Contestability requires a practical opportunity to correct data, present evidence, obtain human review, and receive a reasoned response.

7. Auditability and Model Risk Management

An auditable system preserves evidence of what was known, decided, changed, and executed. Essential artefacts include data dictionaries, provenance, consent and legal-basis records, feature definitions, versioned code, model cards, evaluation results, fairness analyses, approval records, access logs, incident records, and monitoring reports.

A model-risk lifecycle commonly includes:

1. Inventory and classification: identify model purpose, users, affected populations, criticality, and regulatory category.

2. Development: define the target, assumptions, data quality criteria, validation design, and unacceptable uses.

3. Independent validation: test performance, stability, bias, leakage, robustness, explainability, and sensitivity to assumptions.

4. Approval: obtain documented sign-off from accountable business, technical, privacy, security, and risk functions.

5. Deployment controls: restrict access, define human oversight, establish thresholds, and record decisions.

6. Monitoring: track drift, performance, subgroup outcomes, incidents, overrides, and complaints.

7. Change and retirement: revalidate material changes and securely retire data, models, credentials, and interfaces.

A three-lines model is useful: operational teams own and control risk; independent risk, compliance, or validation functions challenge and oversee; internal audit provides independent assurance. Small organisations may combine roles, but they should preserve independence of challenge as far as practicable.

8. Governance Operating Model

Effective governance assigns decision rights rather than creating only committees. A typical structure includes:

• Board or executive risk committee: sets risk appetite and receives material-risk reporting.
• Chief data or analytics officer: establishes enterprise standards and portfolio priorities.
• Data owners: decide permitted use, quality requirements, access, and retention for domains.
• Data stewards: maintain definitions, metadata, quality rules, and issue resolution.
• Model owner: remains accountable for purpose, performance, controls, and lifecycle.
• Privacy officer or data-protection officer: advises on rights, lawful processing, impact assessments, and regulatory engagement.
• Security officer: governs threat assessment, access, resilience, and incident response.
• Independent validation and internal audit: test effectiveness and challenge unsupported claims.
• Affected-person representatives and domain experts: contribute contextual knowledge and identify harms that metrics may miss.

A RACI matrix—responsible, accountable, consulted, informed—should be created for collection, sharing, approval, monitoring, incidents, and deletion. Governance should be risk-tiered: a low-risk internal dashboard may require standard review, whereas an automated employment or credit decision requires enhanced assessment, validation, human oversight, and documented appeal.

9. Responsible Analytics Lifecycle

A practical lifecycle embeds controls at each stage:

Frame: define the legitimate objective, affected parties, foreseeable harms, alternatives, and success criteria beyond accuracy.

Acquire: establish provenance, lawful basis, consent where applicable, data quality, minimisation, and supplier obligations.

Prepare: document transformations, detect leakage, protect sensitive fields, assess representativeness, and preserve lineage.

Develop: compare interpretable and complex models, evaluate subgroup performance, test robustness, and record assumptions.

Decide: conduct privacy and impact assessments, obtain approvals, define human oversight, and communicate limitations.

Deploy: enforce access controls, logging, versioning, monitoring, rate limits, and incident escalation.

Review: monitor drift, fairness, security, complaints, overrides, and changing law; revalidate or withdraw the system when risks exceed tolerance.

10. Worked Quantitative Application: Fairness Diagnosis

Suppose a screening model is evaluated for two groups. For Group A, TP = 72 and FN = 18, so the true-positive rate is 72/(72 + 18) = 0.80. For Group B, TP = 54 and FN = 36, so the true-positive rate is 54/(54 + 36) = 0.60. The model therefore has a 20-percentage-point difference in opportunity for qualified cases.

For false positives, suppose Group A has FP = 20 and TN = 80. Its FPR is:

20/(20 + 80) = 20/100 = 0.20.

Suppose Group B has FP = 10 and TN = 90. Its FPR is 10/(10 + 90) = 0.10. The model produces both lower true-positive and lower false-positive rates for Group B. This does not by itself establish intentional discrimination, because base rates, labels, and selection processes may differ, but it establishes a material disparity requiring investigation. Possible responses include improving labels, collecting representative data, changing thresholds with documented justification, adding human review, and assessing the social cost of each error type. A single aggregate accuracy figure would conceal these differences.

11. Illustrative Case: Hospital Readmission Model

Background

A regional hospital deploys a model to identify patients likely to be readmitted within 30 days. The stated purpose is to offer follow-up support, transport assistance, medication counselling, and earlier clinical review.

Problem

The model uses historical utilisation and billing data. An audit finds that patients from a low-income neighbourhood receive fewer referrals despite comparable clinical need. The data also contain a proxy for ability to pay. Clinicians cannot explain several predictions, and a vendor contract permits broad reuse of patient-level data.

Analysis

The intended intervention is supportive, but the model’s target reflects historical access to care rather than need alone. The proxy may encode structural inequality. The use involves sensitive health data and requires HIPAA analysis, minimum-necessary access, contractual controls, security safeguards, and a documented clinical governance process. The team should examine subgroup FNR and FPR, calibration, missingness, referral access, and whether interventions are actually available equally.

Approach and solution

The hospital limits variables to clinically justified features, removes the payment proxy, and conducts a privacy impact assessment. It renegotiates the vendor contract to prohibit unrelated reuse and require deletion. An independent validation team tests performance and fairness by race, disability, language, age, and intersectional groups. Clinicians receive reason codes and can override predictions with documented justification. Patients are informed that risk scoring supports—not replaces—clinical judgement and can request correction. A monitoring dashboard tracks drift, referrals, outcomes, complaints, and subgroup disparities.

Outcome

The revised model has slightly lower aggregate discrimination but materially reduces the disparity in missed high-risk cases. The hospital retains the system only because the intervention is beneficial, accessible, monitored, and subject to human review. If those conditions cease to hold, governance rules require suspension and reassessment.

Learning points

The case demonstrates that a model’s purpose, intervention, data provenance, procurement terms, fairness definition, and institutional context are inseparable. Technical remediation alone cannot correct an unfair programme if resources are distributed unequally.

12. Challenges and Contemporary Developments

Governance faces several persistent challenges. Data are copied across jurisdictions and suppliers, making lineage difficult. Privacy-enhancing techniques can reduce utility or create false confidence. Fairness criteria may conflict. Explainability may be inadequate for complex models. Generative AI can produce confident errors, expose confidential prompts, reproduce training bias, and create uncertain provenance. Automated monitoring may itself become surveillance. Small organisations may lack specialist staff, while excessive procedural controls can prevent beneficial innovation.

Contemporary practice therefore combines privacy-enhancing computation—such as federated learning, secure multiparty computation, trusted execution environments, and differential privacy—with stronger procurement, model evaluation, red teaming, content provenance, and human-centred design. These methods are not universal solutions. Their suitability depends on threat models, computational assumptions, operational maturity, and the consequences of failure.

Best-Practice Checklist

1. Maintain a complete inventory of datasets, models, owners, purposes, jurisdictions, and suppliers.

2. Conduct proportional privacy, security, ethical, and algorithmic-impact assessments before deployment.

3. Document lawful basis, consent, minimisation, retention, rights, and transfer controls.

4. Use least privilege, encryption, secrets management, secure development, logging, and tested response plans.

5. Evaluate performance and error rates for relevant and intersectional groups.

6. Prefer interpretable approaches in high-impact settings unless complexity is justified by evidence.

7. Provide notice, meaningful human oversight, correction, appeal, and remediation.

8. Independently validate material models and revalidate after significant changes.

9. Monitor drift, misuse, incidents, disparities, complaints, and vendor performance.

10. Treat governance as continual improvement rather than a one-time approval.

Chapter Summary

Responsible analytics integrates ethical principles, privacy, security, legal compliance, fairness, explainability, auditability, and model-risk management. Privacy requires lawful and purposeful collection, minimisation, controlled sharing, rights processes, and defensible retention. Security protects analytical data, pipelines, models, and interfaces through layered preventive and detective controls. GDPR, CCPA/CPRA, HIPAA, the EU AI Act, NIST AI RMF, and ISO standards differ in scope and legal status, so organisations need a mapped, jurisdiction-sensitive obligations register. Fairness requires explicit harm analysis and subgroup evaluation rather than reliance on accuracy or a single metric. Explainability must support understanding and contestability, while auditability depends on durable evidence and independent challenge. A risk-tiered lifecycle with clearly assigned governance roles makes responsible analytics operational and measurable.

Key Terms

Ethics: Principles concerning morally responsible conduct and avoidance of harm.

Privacy: Appropriate control over information relating to identifiable persons.

Information security: Protection of information and systems from unauthorised compromise.

Data governance: Decision rights, responsibilities, standards, and controls for data.

Consent: An informed and meaningful indication of agreement to specified processing.

Purpose limitation: Restriction of data use to defined and compatible purposes.

Data minimisation: Collection and retention of only data necessary for a legitimate purpose.

Pseudonymisation: Replacement of identifiers while retaining a possible linkage mechanism.

Anonymisation: Processing intended to make re-identification reasonably unlikely.

Differential privacy: A formal bound limiting the effect of one individual’s data on an output.

Fairness: The justified and non-arbitrary treatment of people or groups in an analytical process.

False-positive rate: FP divided by FP plus TN.

Explainability: The use of methods and information to clarify model behaviour or outputs.

Contestability: The practical ability to challenge, correct, and obtain review of a decision.

Auditability: The ability to reconstruct and evaluate decisions, changes, and controls from evidence.

Model risk: The possibility of adverse consequences from model errors, misuse, or limitations.

Data steward: A role responsible for definitions, metadata, quality, and operational data controls.

Data-protection impact assessment: A structured assessment of privacy risks and mitigations.

Purpose drift: Expansion or change of data use beyond the originally justified purpose.

Human oversight: Meaningful human authority and capability to review, intervene, or stop a system.

Review Questions

Short Answer Questions

1. Distinguish privacy, security, compliance, and data governance.

2. Why does pseudonymisation not necessarily remove personal-data status?

3. Define purpose limitation and data minimisation.

4. What is the difference between global and local explainability?

5. State the formula for false-positive rate.

Long Answer Questions

1. Compare GDPR, CCPA/CPRA, HIPAA, and the EU AI Act in terms of scope, rights, obligations, and analytical implications.

2. Explain how a model-risk management lifecycle supports accountable analytics.

3. Discuss why fairness metrics can conflict and how an organisation should choose among them.

4. Design a security architecture for a cloud-based analytical platform handling sensitive data.

5. Explain the roles of data owners, stewards, model owners, privacy officers, security officers, validation teams, and internal audit.

Analytical / Application Questions

1. A hiring model has equal accuracy across two groups but substantially higher false-negative rates for one group. Diagnose possible causes and propose a governance response.

2. A retailer wants to reuse loyalty-card data, collected for discounts, to estimate insurance risk. Analyse the purpose-limitation, consent, fairness, security, and compliance issues.

3. Design a risk-tiered approval and monitoring plan for a hospital model that recommends patients for additional care support.

Multiple Choice Questions with Answers

Question 1

A multinational organisation is deploying an automated employment-screening system affecting EU residents. Which combination most directly reflects a defensible governance approach?

A. Rely only on aggregate accuracy because employment decisions are business decisions

B. Map GDPR obligations and applicable AI requirements, conduct impact assessment, validate subgroup outcomes, and provide human review

C. Remove the protected attribute and treat the system as unbiased

D. Use a vendor contract as a substitute for organisational accountability

Correct Answer: B. Map GDPR obligations and applicable AI requirements, conduct impact assessment, validate subgroup outcomes, and provide human review

Question 2

A dataset has names removed but retains rare dates, locations, and occupations, while a separate table permits linkage. What is the most accurate classification?

A. It is necessarily anonymous because direct identifiers were removed

B. It is pseudonymised or otherwise potentially personal data because linkage and re-identification remain possible

C. It is encrypted data and therefore outside privacy regulation

D. It is automatically compliant if used for research

Correct Answer: B. It is pseudonymised or otherwise potentially personal data because linkage and re-identification remain possible

Question 3

For a high-impact credit model, which control most directly supports contestability?

A. Increasing model complexity to improve predictive accuracy

B. Publishing only the model’s aggregate accuracy

C. Providing relevant reasons, correction procedures, and meaningful human review

D. Deleting all decision logs after deployment

Correct Answer: C. Providing relevant reasons, correction procedures, and meaningful human review

Question 4

A classifier produces FP = 20 and TN = 80 for a group. What is its false-positive rate?

A. 0.20

B. 0.25

C. 0.80

D. 0.05

Correct Answer: A. 0.20

Question 5

Two groups have different base rates. An organisation observes that equalised odds and equal positive predictive value cannot both be achieved under the current score distributions. What is the most appropriate conclusion?

A. One metric must have been calculated incorrectly

B. The organisation should select the metric without considering the harm or decision context

C. Fairness criteria may conflict, so the organisation must justify the relevant harm, metric, and trade-off

D. Removing the protected attribute will guarantee both conditions

Correct Answer: C. Fairness criteria may conflict, so the organisation must justify the relevant harm, metric, and trade-off

Question 6

A hospital model uses a variable strongly correlated with ability to pay, although that variable is not a protected attribute. What is the principal governance concern?

A. The variable cannot affect fairness because it is not legally protected by name

B. It may act as a proxy and reproduce structural disparities, requiring justification and impact analysis

C. It proves that the model is causally valid

D. It eliminates the need for subgroup testing

Correct Answer: B. It may act as a proxy and reproduce structural disparities, requiring justification and impact analysis

Question 7

Which allocation best preserves independence in a three-lines model?

A. The model developer approves, validates, and audits the same model

B. Operational teams own risk, an independent risk function challenges it, and internal audit provides independent assurance

C. Internal audit builds the model and owns its business outcome

D. The vendor alone determines acceptable model risk

Correct Answer: B. Operational teams own risk, an independent risk function challenges it, and internal audit provides independent assurance

Question 8

A model’s input distribution has shifted substantially after deployment, but outcome labels will not be available for several months. What is the most defensible immediate action?

A. Ignore the shift until accuracy can be calculated

B. Treat the shift as a monitoring signal, investigate data and population changes, apply safeguards, and consider temporary restrictions

C. Delete historical monitoring records

D. Increase the decision threshold without documenting the reason

Correct Answer: B. Treat the shift as a monitoring signal, investigate data and population changes, apply safeguards, and consider temporary restrictions

Question 9

Which statement best distinguishes NIST AI RMF and the EU AI Act?

A. Both are identical criminal statutes

B. NIST AI RMF is a voluntary risk-management framework, whereas the EU AI Act is a legal regulatory regime with risk-based obligations

C. The EU AI Act applies only to cybersecurity, while NIST AI RMF applies only to privacy

D. NIST AI RMF replaces all national and sectoral law

Correct Answer: B. NIST AI RMF is a voluntary risk-management framework, whereas the EU AI Act is a legal regulatory regime with risk-based obligations

Question 10

A vendor offers a predictive model but refuses to provide version history, training-data provenance, performance by subgroup, or incident information. What should the procuring organisation do first?

A. Deploy immediately because vendor responsibility removes buyer responsibility

B. Accept the model if its marketing accuracy is high

C. Treat the missing evidence as a material third-party risk and require documentation, validation rights, and contractual controls before approval

D. Avoid all documentation because it could expose trade secrets

Correct Answer: C. Treat the missing evidence as a material third-party risk and require documentation, validation rights, and contractual controls before approval