Ethics, Risk and Machine-Learning Governance

Introduction

Machine learning has moved from experimental analytics into decisions that affect credit, employment, insurance, healthcare, education, public services and customer relationships. This expansion creates strategic value, but it also creates obligations. A model may be statistically accurate and still be unacceptable if it discriminates, exposes personal information, cannot be explained, is vulnerable to manipulation, or operates without meaningful accountability.

Machine-learning governance is the system of policies, roles, processes, technical controls and evidence used to ensure that models are lawful, fair, secure, reliable, transparent and aligned with organizational purposes. Governance is broader than model validation. It includes the complete lifecycle: problem definition, data collection, feature engineering, development, testing, deployment, monitoring, change management, retirement and post-incident review.

For business analytics professionals, ethics is not an optional layer added after technical development. It affects which problem should be solved, which data may be used, what performance means, who bears the risk and whether a model should be deployed at all.

Learning Objectives

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

1. Assess fairness, privacy, transparency, security, accountability and environmental impacts of machine-learning systems.
2. Distinguish data bias from algorithmic, measurement and deployment bias.
3. Apply appropriate fairness metrics while recognizing their assumptions and trade-offs.
4. Establish documentation, validation, human-oversight and audit controls.
5. Relate organizational practice to major responsible-AI and regulatory frameworks.
6. Design a proportionate governance process for a high-impact analytics application.
7. Critically evaluate the limitations of technical solutions to social and ethical problems.

1. From Model Performance to Responsible Decision-Making

Traditional model assessment emphasizes predictive performance. For a binary classifier, a confusion matrix distinguishes true positives, false positives, true negatives and false negatives. Common measures include accuracy, precision, recall and the false-positive rate.

Here TP, TN, FP and FN denote true positives, true negatives, false positives and false negatives. The expression assumes a binary classification task and a defined decision threshold. Accuracy can be misleading when classes are imbalanced or when the costs of errors differ.

Precision is the proportion of predicted positives that are actually positive. It is undefined when TP + FP = 0.

Recall, also called sensitivity or the true-positive rate, is the proportion of actual positives detected by the model. It is undefined when TP + FN = 0.

A responsible system adds questions such as:

• Is the target variable a legitimate and suitable representation of the decision objective?
• Are errors distributed unequally across groups?
• Were individuals informed that their data would be used?
• Can affected people challenge or correct an outcome?
• Is the model secure against unauthorized access or manipulation?
• Is there a named person or committee accountable for decisions?
• Are the benefits proportionate to social, legal and environmental costs?

These questions may require a decision not to automate. Governance therefore concerns legitimacy and institutional responsibility, not merely optimization.

2. Ethical Risk Dimensions

2.1 Fairness and discrimination

Fairness concerns whether people and groups are treated justly. It is not a single mathematical property. Relevant dimensions include equality of treatment, equality of opportunity, equality of outcomes, individual consistency, procedural fairness and the avoidance of unjustified disadvantage.

Protected characteristics may include race, ethnicity, sex, gender identity, disability, age, religion or other legally protected categories. Even when such variables are excluded, proxy variables—such as postcode, language, purchasing behavior or educational history—may reproduce protected characteristics.

A useful distinction is between:

• Direct discrimination: a decision explicitly uses a protected characteristic in an unjustified way.
• Indirect discrimination: a neutral-looking rule disproportionately disadvantages a protected group.
• Measurement bias: the target or features measure some groups less accurately.
• Representation bias: training data do not adequately represent the deployment population.
• Historical bias: data reflect earlier discriminatory institutions or practices.
• Aggregation bias: one model is applied to groups with different relationships between features and outcomes.
• Evaluation bias: performance is measured using benchmarks that favor some populations.
• Deployment bias: a technically valid model is used in a context or manner different from that assumed during development.

The distinction between historical patterns and legitimate prediction is especially important. A model that predicts past hiring decisions may reproduce past exclusion rather than identify job capability.

2.2 Privacy and data protection

Privacy includes control over personal information, appropriate use, confidentiality, freedom from unjustified surveillance and protection against inference. Data protection laws commonly require principles such as lawfulness, fairness, transparency, purpose limitation, data minimization, accuracy, storage limitation, integrity, confidentiality and accountability. Exact obligations depend on jurisdiction, sector and organizational role.

Important privacy risks include:

• collection of data beyond the stated purpose;
• re-identification of supposedly anonymous records;
• inference of sensitive attributes;
• unauthorized secondary use;
• excessive retention;
• insecure data sharing;
• model inversion, in which information about training records is inferred;
• membership inference, in which an attacker estimates whether a person was in the training set.

Anonymization seeks to make re-identification reasonably infeasible, whereas pseudonymization replaces direct identifiers but does not eliminate personal-data status in many legal regimes. Small datasets, rare combinations of attributes and external data can make re-identification easier than expected.

Privacy-preserving techniques include access controls, encryption, data minimization, aggregation, retention limits, federated learning and differential privacy. Differential privacy provides a formal privacy guarantee by limiting the effect of any single record on an output. For a randomized mechanism M, ε-differential privacy requires that for neighboring datasets D and D′ differing in one individual, and for every output set S:

Here P is probability, ε ≥ 0 is the privacy-loss parameter, and S is any possible set of outputs. Smaller ε generally indicates stronger privacy but often reduces statistical utility. The definition assumes a specified neighboring-dataset relationship and a correctly implemented randomized mechanism; it does not guarantee that every use is ethically acceptable.

2.3 Transparency and explainability

Transparency concerns whether relevant information about a system is available. Explainability concerns whether its operation or outputs can be meaningfully understood by appropriate audiences. These are related but distinct.

Different stakeholders need different explanations:

• a data subject may need the main factors influencing a decision and a route to challenge it;
• a model developer may need feature behavior, error analysis and reproducibility;
• an auditor may need data lineage, approvals, tests and change records;
• an executive may need risk, benefit, cost and accountability information;
• a regulator may need evidence that legal obligations were met.

A simple model is not automatically transparent, and a complex model is not automatically unacceptable. Explanations can be misleading when they are post hoc approximations, describe average rather than individual behavior, or imply causal relationships that the model has not established. Explanations should therefore be tested for fidelity, stability, comprehensibility and usefulness.

2.4 Security and adversarial risk

Machine-learning systems enlarge the attack surface of an organization. Threats may target data, models, interfaces, infrastructure or human operators. Examples include:

• data poisoning, in which training examples are manipulated;
• evasion, in which inputs are altered to cause incorrect predictions;
• model theft through repeated queries;
• prompt or input manipulation in systems using language models;
• unauthorized access to training data;
• supply-chain compromise through third-party models or packages;
• denial of service;
• leakage through logs, explanations or error messages.

Security controls should include threat modeling, identity and access management, encryption, secrets management, dependency scanning, secure development practices, rate limiting, input validation, monitoring and incident response. Security is not equivalent to fairness: a secure model can still discriminate, and a fair model can still be vulnerable.

2.5 Accountability and human agency

Accountability means that responsibilities are assigned, decisions can be reconstructed and remedies are available. A statement that “the algorithm decided” is not an adequate governance arrangement. Organizations should identify the business owner, model owner, data owner, technical approver, risk or compliance reviewer and operational decision-maker.

Human oversight must be meaningful rather than ceremonial. A reviewer who lacks time, information, authority or expertise may simply rubber-stamp automated recommendations. Effective oversight specifies when humans must review, what evidence they receive, when they may override the system and how overrides are recorded and analyzed.

2.6 Environmental and social sustainability

Training and operating models consume electricity, water and hardware resources. Environmental assessment should consider development experiments, data storage, inference volume, hardware manufacture, model retirement and the consequences of erroneous decisions. A large model is not necessarily the most responsible choice. Smaller models, efficient architectures, reuse of pretrained models, workload scheduling, carbon-aware computing and explicit value assessment may reduce impact.

Environmental reporting must be cautious: emissions depend on hardware, data-center energy sources, utilization, geographic location and accounting boundaries. Reported estimates should state the measurement method and system boundary rather than imply false precision.

3. Measuring and Diagnosing Fairness

Let A denote a group attribute, Ŷ the predicted outcome, Y the observed outcome and X the features used by the model. Fairness analysis begins by identifying the decision, affected groups, relevant harm and legitimate outcome. Metrics should not be selected before these questions are answered.

3.1 Group fairness metrics

Demographic parity requires similar positive prediction rates across groups:

This may be appropriate when equal selection rates are the objective, but it can be unsuitable when groups have different legitimate base rates or when the decision concerns a genuinely different need.

Equal opportunity requires equal true-positive rates:

This focuses on access to correctly identified positive cases. It is useful when false negatives are the principal concern.

Equalized odds requires both true-positive and false-positive rates to be equal across groups:

Predictive parity requires equal positive predictive value:

Calibration within groups requires that predicted probabilities have the same interpretation across groups. For example, among people assigned a predicted risk of 0.70, the observed event rate should be approximately 0.70 in each group, subject to sampling uncertainty and calibration methodology.

These criteria can conflict. When groups have different base rates, nontrivial prediction error exists and group membership affects risk distributions, it is generally impossible to satisfy calibration and equalized error criteria simultaneously. This is not a technical defect that one universal metric can solve; it is a policy choice requiring explicit discussion of harms, benefits and legal requirements.

3.2 A worked fairness example

Suppose a classifier is used to identify applicants for an intervention. The following results are observed:

For Group A:

For Group B:

The true-positive-rate difference is 0.05, or 5 percentage points. The false-positive rates are:

The results suggest a disparity in both error rates. However, they do not by themselves establish unlawful discrimination. Analysts must investigate sample uncertainty, data quality, legitimate differences in context, threshold selection, the definition of Y and the consequences of each error. Reporting confidence intervals or statistical tests may be appropriate, but statistical significance does not determine ethical significance.

3.3 Beyond group averages

Group metrics can conceal intersectional disadvantage. A model may appear acceptable for women and for ethnic minorities separately while performing poorly for women in a particular ethnic minority. Analysts should inspect intersectional groups where sample size permits, examine distributions rather than only averages, and assess individual cases.

Fairness is also longitudinal. A model may impose repeated burdens on the same people, create feedback loops or change behavior in response to prediction. Monitoring should therefore examine outcomes over time, not only a pre-deployment test set.

4. Governance Across the Machine-Learning Lifecycle

A mature governance process uses proportionate controls. High-impact systems require stronger evidence, independent review and more frequent monitoring than low-risk recommendation tools.

4.1 Problem formulation and impact assessment

Before collecting data, document:

• the decision and its legitimate purpose;
• affected individuals and groups;
• expected benefits and possible harms;
• whether automation is necessary;
• prohibited uses;
• applicable law, contracts and sector rules;
• human alternatives and appeal routes;
• success criteria beyond predictive accuracy.

An algorithmic impact assessment should consider foreseeable misuse, vulnerable populations, power asymmetries and whether the organization can provide remedy. The appropriate output may be redesign, restricted deployment or non-deployment.

4.2 Data governance

Data governance should establish provenance, ownership, lawful basis, consent or other authorization where relevant, quality controls, retention periods, access permissions and permitted uses. A data card or dataset register should record collection context, population coverage, known exclusions, labeling procedures, missingness, sensitive attributes, transformations and limitations.

Train–test separation remains essential for valid evaluation, but it is not sufficient for governance. Leakage, temporal shifts and selective labels can produce deceptively strong results. Validation data should reflect the intended deployment population and operating conditions.

4.3 Model development and validation

Model documentation should include the objective, intended use, out-of-scope use, features, target definition, training data, assumptions, performance by relevant subgroup, calibration, robustness, security testing, explainability method and known limitations.

Independent validation should challenge:

• the appropriateness of the target;
• data lineage and leakage controls;
• reproducibility;
• performance and uncertainty;
• fairness and accessibility;
• privacy and security;
• operational controls;
• compliance with approval conditions.

A model should not be approved solely because it exceeds a benchmark. Validation should assess whether the benchmark represents the actual decision and whether improvements justify added complexity and risk.

4.4 Deployment, monitoring and change control

A deployment plan should define access privileges, interfaces, fallback procedures, human review, logging, alert thresholds and incident ownership. Monitoring should cover data drift, concept drift, performance, calibration, subgroup error, missingness, latency, security events, override rates and complaints.

A useful conceptual distinction is:

• Data drift: the distribution of inputs changes.
• Concept drift: the relationship between inputs and outcomes changes.
• Performance degradation: predictive quality falls.
• Governance drift: actual use diverges from approved use.

Change control should classify updates by risk. A new feature, threshold, training dataset, vendor model or decision context may require partial or complete revalidation. Model retirement should be planned because obsolete models can remain embedded in workflows.

4.5 Documentation and auditability

A model inventory should identify every production and material experimental model, its owner, risk classification, version, status, data sources, dependencies and review date. Useful artifacts include:

• dataset documentation;
• model cards;
• decision records;
• impact assessments;
• validation reports;
• approval minutes;
• audit logs;
• incident records;
• user training records;
• monitoring dashboards;
• decommissioning records.

Documentation is not merely administrative. It creates institutional memory, supports reproducibility and allows an auditor to reconstruct what was known and authorized at a particular time.

5. Regulatory and Responsible-AI Frameworks

Organizations should obtain jurisdiction-specific legal advice; frameworks do not replace law. Nevertheless, several influential approaches provide useful governance structures.

The OECD AI Principles emphasize inclusive growth, sustainable development, human-centered values, transparency, robustness, security, safety and accountability. The UNESCO Recommendation on the Ethics of Artificial Intelligence emphasizes human rights, environmental sustainability, diversity, fairness, transparency, human oversight and impact assessment.

The NIST AI Risk Management Framework organizes practice around governing, mapping, measuring and managing risks. It is voluntary in many settings but influential because it connects organizational governance with technical measurement and operational response.

ISO/IEC 23894 addresses AI risk management, while ISO/IEC 42001 specifies requirements for an AI management system. ISO/IEC 27001 remains relevant to information-security management, and ISO/IEC 27701 can support privacy-information management.

The EU AI Act adopts a risk-based structure, including prohibited practices, high-risk systems, transparency obligations for certain systems and obligations for providers and deployers. Its requirements depend on system classification and role. The General Data Protection Regulation may apply to personal-data processing and includes principles concerning lawfulness, transparency, data minimization, rights and automated decision-making in specified circumstances. Other jurisdictions have adopted or proposed privacy, automated-decision, anti-discrimination and sector-specific rules.

A responsible organization maps legal obligations and voluntary controls to concrete evidence. For example, a transparency requirement should correspond to notices, explanation procedures and records; an accountability requirement should correspond to named owners, approval evidence and audit trails.

6. An Integrated Governance Operating Model

Governance is effective when responsibilities are distributed but coordinated. A three-lines model is useful:

1. First line—business and technical owners: develop, operate and monitor systems; manage day-to-day controls.
2. Second line—risk, legal, privacy and compliance: set standards, challenge decisions and advise on obligations.
3. Third line—internal audit or independent assurance: evaluate whether governance and controls operate effectively.

A model risk committee may classify systems according to impact, autonomy, data sensitivity, affected population, reversibility of decisions and potential harm. Risk classification should determine approval authority, validation depth, monitoring frequency and incident escalation.

A practical control cycle is:

1. Identify purpose, stakeholders and harms.
2. Classify risk and applicable obligations.
3. Document data and intended use.
4. Develop with privacy, fairness and security controls.
5. Validate independently.
6. Approve with explicit conditions.
7. Deploy with human oversight and logging.
8. Monitor outcomes and complaints.
9. Investigate incidents and apply remedies.
10. Revalidate, restrict or retire the system.

This cycle should include affected-community consultation where decisions materially affect people who may not be represented within the organization.

7. Detailed Applied Case: Automated Small-Business Lending

The following is an illustrative case, not a report of a particular organization.

Background

A financial institution proposes a machine-learning system to prioritize small-business loan applications. The model uses transaction history, cash-flow patterns, business age, industry, location and existing banking relationships. It produces a risk score that informs underwriting.

Problem

During pilot testing, approval rates are higher for established urban firms than for newer firms in rural areas. The model excludes race and sex, but location and transaction characteristics may act as proxies. Historical loan outcomes also reflect earlier underwriting policies, so the target “repaid loan” may combine borrower capacity with prior access to favorable credit.

Analysis

The institution defines the decision, affected groups and harms. It discovers that rural applicants have more missing digital transactions because some payments occur outside the institution’s platform. Missingness is therefore not random. The model interprets limited transaction history as elevated risk.

Validation reports show:

• high overall discrimination performance;
• lower recall for rural applicants;
• higher false-positive rates for newer firms;
• poorer calibration in low-volume industries;
• no documented appeal process;
• insufficient evidence that the model’s explanations are understandable to applicants.

The organization also identifies a privacy concern: retaining detailed transaction data indefinitely would exceed the operational need.

Approach and solution

The institution takes several actions:

1. It revises the target and separates repayment capacity from historical approval decisions where feasible.
2. It adds a data-quality indicator and investigates whether missingness reflects access conditions rather than risk.
3. It evaluates subgroup recall, false-positive rates, calibration and intersectional groups.
4. It tests alternative thresholds and a policy that sends uncertain cases to trained human underwriters.
5. It removes features that provide little incremental value but create substantial proxy risk.
6. It establishes a documented appeal and correction process.
7. It limits retention, encrypts sensitive data and restricts access by role.
8. It creates a model card, data register, validation report and approval record.
9. It monitors approval, pricing, override, complaint and repayment outcomes by subgroup.
10. It schedules independent review before material model changes.

Outcome

The final system has slightly lower aggregate predictive performance but materially better subgroup performance and clearer operational safeguards. The institution restricts automated adverse decisions and requires human review for low-confidence or materially adverse cases. It also discovers that some apparently “model” problems originate in unequal access to digital banking data. A non-model intervention—improving data access and application support—becomes part of the solution.

Learning points

The case demonstrates that fairness cannot be inferred from excluding protected attributes. Target construction, measurement quality, historical decisions and deployment procedures matter. It also shows why governance must combine technical tests with process changes, human remedy and organizational accountability.

8. Challenges and Limitations

Technical fairness metrics depend on labels that may themselves be biased or incomplete. A metric can be satisfied while people experience unfair treatment in a broader process. Removing sensitive attributes can make diagnosis harder and can prevent appropriate fairness correction; retaining them requires strong safeguards and a legitimate purpose.

Explainability methods may provide plausible but inaccurate narratives. Privacy techniques can reduce utility, and privacy risk can remain through auxiliary information. Human review can reduce automation bias but can also introduce inconsistency, delay and discretionary discrimination. Independent audits can improve assurance but may be limited by restricted access, weak standards, vendor secrecy or insufficient expertise.

Regulation is fragmented across jurisdictions and changes over time. Voluntary principles may be interpreted inconsistently. Small organizations may lack specialist resources, while large organizations may create extensive documentation without changing decisions. Governance should therefore emphasize evidence of effective control rather than paperwork alone.

Finally, not every harm can be solved through model adjustment. Some problems arise from institutional inequality, insufficient services, poor measurement or an inappropriate objective. Responsible analytics sometimes requires redesigning the process or declining to automate.

9. Best-Practice Checklist

A graduate-level practitioner should be able to ask:

• Is the use necessary, proportionate and within a legitimate purpose?
• Who is affected, and who may be missing from the data or consultation?
• What harms arise from false positives, false negatives and systematic exclusion?
• Are protected and proxy attributes understood and appropriately governed?
• Is the target a valid measure of the intended outcome?
• Are subgroup and intersectional performance results documented?
• Are privacy, security and environmental impacts assessed?
• Can people understand, challenge and correct important decisions?
• Are owners, approvers and escalation routes named?
• Is there independent validation and an audit trail?
• Are monitoring, incident response, change control and retirement defined?
• Would a reasonable affected person regard the process as legitimate?

Chapter Summary

Machine-learning governance extends beyond predictive accuracy to include fairness, privacy, transparency, security, accountability, sustainability and human agency. Bias may enter through historical decisions, unrepresentative samples, measurement choices, proxies, aggregation and deployment practices. Fairness metrics such as demographic parity, equal opportunity, equalized odds, predictive parity and calibration illuminate different concerns and may conflict; they must therefore be selected in relation to the decision, harm and legal context.

Responsible governance covers the full lifecycle. It begins with impact assessment and legitimate problem formulation, continues through data and model documentation, independent validation, secure deployment, human oversight, monitoring, incident response, change control and retirement. Model cards, dataset documentation, audit logs and named owners create evidence for accountability. Frameworks such as the OECD AI Principles, UNESCO Recommendation, NIST AI RMF, ISO/IEC standards, the EU AI Act and data-protection law provide useful reference points, but organizations must translate them into operational controls.

The central strategic lesson is that responsible AI is a sociotechnical discipline. Technical interventions are necessary but insufficient. Effective practice combines statistical analysis, legal and ethical reasoning, organizational design, stakeholder participation and a willingness to redesign or reject automation when risks cannot be justified.

Key Terms

Fairness: The principle that people and groups should be treated justly and without unjustified disadvantage.

Algorithmic bias: Systematic error or disadvantage produced or amplified by an algorithmic system.

Historical bias: Bias embedded in data because past institutions or decisions were unequal.

Proxy variable: A feature that indirectly represents a protected or sensitive attribute.

Demographic parity: Similar positive prediction rates across groups.

Equal opportunity: Similar true-positive rates across groups.

Equalized odds: Similar true-positive and false-positive rates across groups.

Calibration: Agreement between predicted probabilities and observed frequencies.

Privacy: Protection of personal information, autonomy and appropriate control over data use.

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

Explainability: The extent to which model behavior or outputs can be understood.

Accountability: Assignment of responsibility, traceability and availability of remedy.

Human oversight: Meaningful human capacity to review, challenge, override or stop system operation.

Data drift: Change in the distribution of model inputs over time.

Concept drift: Change in the relationship between inputs and outcomes.

Model card: Documentation describing a model’s purpose, performance, limitations and uses.

Algorithmic impact assessment: A structured evaluation of potential effects, risks and affected stakeholders.

Model risk: The possibility of adverse consequences caused by incorrect, misused or poorly governed models.

Responsible AI: The design and operation of AI systems consistent with ethical, legal, social and organizational responsibilities.

Review Questions

Short Answer Questions

1. Why is predictive accuracy insufficient for responsible machine-learning deployment?
2. Distinguish historical bias, measurement bias and deployment bias.
3. What does equal opportunity measure?
4. Why is pseudonymization not necessarily full anonymization?
5. Name three artifacts that support model auditability.

Long Answer Questions

1. Explain how fairness metrics can conflict when groups have different base rates. What should an organization do when no single metric can satisfy all objectives?
2. Describe a complete machine-learning governance lifecycle from problem formulation to model retirement.
3. Compare transparency, explainability and accountability. Why are these concepts not interchangeable?
4. Discuss the privacy and security risks of using detailed personal data in machine-learning systems and propose controls.
5. Evaluate the role of NIST, ISO, OECD, UNESCO and regulatory frameworks in responsible-AI governance.

Analytical / Application Questions

1. A hiring model has equal overall accuracy for two groups but substantially higher false-negative rates for one group. Diagnose possible causes and design a validation and remediation plan.
2. Design a proportionate governance process for a healthcare triage model that affects access to treatment. Specify owners, evidence, human oversight, monitoring and escalation controls.
3. An organization claims that its model is fair because it excludes race and sex. Critically assess the claim using proxies, target construction, intersectionality and deployment context.

Multiple Choice Questions with Answers

Question 1

Which fairness criterion requires equal true-positive rates across groups?

A. Demographic parity

B. Equal opportunity

C. Predictive parity

D. Calibration

Correct Answer: B. Equal opportunity

Question 2

What is the principal purpose of a model card?

A. To encrypt model parameters

B. To replace independent validation

C. To document a model’s purpose, performance and limitations

D. To guarantee legal compliance

Correct Answer: C. To document a model’s purpose, performance and limitations

Question 3

Which example best illustrates measurement bias?

A. A training sample excludes rural users

B. A feature records medical symptoms less accurately for one group

C. An attacker alters an input at prediction time

D. A model is deployed after its approved use has changed

Correct Answer: B. A feature records medical symptoms less accurately for one group

Question 4

Under differential privacy, a smaller ε generally indicates:

A. Stronger privacy protection

B. Higher model accuracy in every setting

C. No need for access controls

D. Removal of all statistical noise

Correct Answer: A. Stronger privacy protection

Question 5

Which governance activity should occur before model development?

A. Retiring the model

B. Defining purpose, affected stakeholders and potential harms

C. Calculating post-deployment drift

D. Deleting all validation records

Correct Answer: B. Defining purpose, affected stakeholders and potential harms

Question 6

A change in the relationship between features and the outcome is called:

A. Data drift

B. Concept drift

C. Representation parity

D. Pseudonymization

Correct Answer: B. Concept drift

Question 7

Why can excluding protected attributes fail to eliminate discrimination?

A. Models cannot operate without protected attributes

B. Proxy variables may reproduce protected characteristics

C. Accuracy automatically becomes zero

D. Privacy law always requires protected attributes

Correct Answer: B. Proxy variables may reproduce protected characteristics

Question 8

Which is an example of a security attack on a machine-learning system?

A. Calibration

B. Data minimization

C. Data poisoning

D. Human review

Correct Answer: C. Data poisoning

Question 9

In the illustrative lending case, why was missing transaction data ethically important?

A. It proved that all rural applicants were fraudulent

B. It could reflect unequal access to digital banking rather than borrower risk

C. It eliminated the need for subgroup testing

D. It guaranteed demographic parity

Correct Answer: B. It could reflect unequal access to digital banking rather than borrower risk

Question 10

Which statement best describes meaningful human oversight?

A. A human automatically accepts every model recommendation

B. A human has authority, information and procedures to review or override decisions

C. A human signs a policy without seeing model outputs

D. A human is involved only after the model is retired

Correct Answer: B. A human has authority, information and procedures to review or override decisions