Responsible Machine Learning and Governance

Introduction

Machine learning (ML) systems increasingly influence lending, recruitment, insurance, healthcare access, fraud detection, pricing, customer service, and public administration. Their value depends not only on predictive accuracy but also on whether their decisions are lawful, fair, secure, explainable, and accountable. A model can be statistically accurate and still produce unacceptable outcomes if it systematically disadvantages a protected group, exposes confidential information, cannot be audited, or is used beyond the purpose for which its data were collected.

Responsible machine learning is therefore an organizational capability rather than a final technical test. It connects data management, statistical evaluation, software engineering, information security, legal compliance, business ethics, and governance. Governance establishes who may make decisions, what evidence is required, how risks are classified, how models are monitored, and what happens when a system fails.

This chapter presents a practical framework for evaluating fairness, privacy, security, transparency, accountability, and regulatory risk. It also explains risk classification, validation, audit trails, model cards, human review, and responsible data management.

Learning Objectives

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

1. Explain the principal dimensions of responsible machine learning.
2. Distinguish fairness concepts and select suitable evaluation measures.
3. Identify privacy, security, transparency, accountability, and regulatory risks.
4. Apply a risk-based governance lifecycle to an ML system.
5. Design validation, documentation, monitoring, human-review, and incident-response controls.
6. Analyze a business case and recommend proportionate governance mechanisms.

1. Why Responsible Machine Learning Matters

Traditional business software generally follows explicit rules. ML systems infer patterns from historical data, which creates several distinctive risks:

• Historical bias: past decisions may reflect discrimination or unequal opportunity.
• Data representation risk: a sample may omit groups, regions, languages, or unusual cases.
• Proxy discrimination: apparently neutral variables, such as postcode, may correlate strongly with protected characteristics.
• Opacity: complex models may be difficult for affected people and decision-makers to understand.
• Distribution shift: economic, social, or technological changes can make training data obsolete.
• Automation bias: employees may accept model outputs without sufficient scrutiny.
• Feedback loops: predictions can alter the environment that later supplies training data.

Responsible governance asks five linked questions:

1. Should the organization use ML for this decision?
2. Can the data and model be used lawfully and ethically?
3. How will performance and impacts be measured?
4. Who is accountable for decisions and failures?
5. What safeguards, review processes, and remedies are available?

Accuracy is important, but it is only one dimension of system quality. A useful objective is to optimize business performance subject to legal, ethical, security, and operational constraints. These constraints are not optional additions; in high-impact contexts they define acceptable system behavior.

2. A Risk-Based Governance Model

Governance should be proportionate to potential harm. A recommendation engine for movie preferences does not normally require the same controls as a system that determines loan eligibility or access to essential services.

A practical classification uses four dimensions:

Organizations may assign a risk score using a documented rubric. For example, each dimension can be rated from 1 to 5, and a weighted score can be calculated as:

Here, R is the overall risk score; I, A, S, and V represent impact, autonomy, scale, and irreversibility ratings; and the weights w_I, w_A, w_S, and w_V are non-negative organizational weights that sum to 1. This is a governance prioritization tool, not an objective measure of moral harm. A high score should trigger stronger review, not replace professional judgment.

A four-tier scheme is often useful:

• Tier 1—Minimal risk: low-impact personalization or operational forecasting.
• Tier 2—Moderate risk: systems that influence employees, customers, or resource allocation but do not independently determine significant outcomes.
• Tier 3—High risk: systems affecting credit, employment, insurance, education, healthcare, safety, or access to essential services.
• Tier 4—Prohibited or unacceptable use: applications that violate law, fundamental rights, or organizational policy, such as intentionally deceptive manipulation or unjustified discriminatory profiling.

Risk classification must be revisited when the use case, population, data, model, or level of automation changes.

3. Fairness and Non-Discrimination

3.1 Fairness is a family of concepts

Fairness cannot be reduced to a single universal metric. Different definitions answer different questions and may conflict when groups have different base rates or when prediction errors differ. The appropriate definition depends on the decision, law, social context, and harm being considered.

Important concepts include:

• Individual fairness: similar individuals should receive similar treatment, requiring a meaningful definition of similarity.
• Group fairness: error rates, selection rates, or benefits should be compared across demographic groups.
• Procedural fairness: people should receive consistent procedures, notice, explanation, and an opportunity to challenge a decision.
• Allocative fairness: benefits and burdens should be distributed justifiably.
• Disparate impact: a neutral-looking practice produces substantially worse outcomes for a protected group.

Let a binary model produce a positive decision for an individual, with protected group [[MATH=g]]. The selection rate for group [[MATH=g]] is:

Here, [[MATH=Ŷ]] is the predicted decision, [[MATH=G]] is group membership, and [[MATH=P]] denotes probability. A selection-rate ratio comparing group [[MATH=a]] with group [[MATH=b]] is:

This ratio is sometimes called a disparate-impact ratio. It is descriptive, not conclusive: a difference may arise from legitimate factors, measurement problems, historical inequality, or discrimination. Legal thresholds vary by jurisdiction and context; a commonly discussed U.S. employment screening rule of thumb is 0.80, but it is not a universal safe harbor.

For labeled outcomes, the true-positive rate and false-positive rate are:

[[MATH=Y]] is the observed outcome. Equal opportunity seeks similar [[MATH=TPR_g]] across groups, while equalized odds seeks similarity in both true-positive and false-positive rates. Calibration within groups asks whether predicted probabilities have comparable meanings across groups; for a score [[MATH=S]], it requires that [[MATH=P(Y=1 | S=s,G=g)]] be similar across groups for relevant score values.

These criteria may be mathematically incompatible when groups have different outcome base rates and predictions are not perfect. Consequently, fairness analysis requires an explicit choice of objective rather than a search for a metric that automatically proves fairness.

3.2 Fairness evaluation process

A responsible evaluation should:

1. Define the decision, affected population, protected characteristics, and potential harms.
2. Examine representation, missingness, label quality, and historical process bias.
3. Compare outcomes and error rates across relevant groups and intersections, such as age and gender together.
4. Test whether features act as proxies for protected characteristics.
5. Analyze thresholds, reject rates, appeals, and human overrides.
6. Consider causal context: equal prediction rates do not necessarily mean equal opportunity.
7. Document trade-offs, remediation, and residual risk.

Possible interventions include improving data collection, removing leakage, redesigning labels, reweighting observations, changing decision thresholds where lawful and justified, using constrained optimization, adding human review, or deciding not to automate the decision. Fairness testing should continue after deployment because population composition and relationships among variables can change.

4. Privacy and Responsible Data Management

Privacy concerns the appropriate collection, use, disclosure, retention, and deletion of information about identifiable or potentially identifiable people. Responsible data management begins before model training.

4.1 Data governance controls

A data inventory should record data sources, owners, subjects, sensitivity, purpose, legal basis, retention period, access permissions, quality limitations, and downstream uses. Core principles include:

• Purpose limitation: collect and use data for specified, legitimate purposes.
• Data minimization: use only information necessary for the intended task.
• Accuracy: correct or annotate inaccurate data and allow appropriate updating.
• Retention limitation: delete or de-identify data when no longer needed.
• Security: protect data against unauthorized access, alteration, loss, and disclosure.
• Accountability: retain evidence that controls were designed and followed.

De-identification reduces risk but does not automatically make data anonymous. Combining supposedly anonymous records with external information may permit re-identification. Pseudonymization replaces direct identifiers with codes but leaves re-identification possible through additional information.

4.2 Privacy-preserving techniques

• Access control and encryption: restrict authorized use and protect data in transit and at rest.
• Aggregation: report group-level results where individual-level detail is unnecessary.
• Differential privacy: add carefully calibrated randomness to limit the effect of any one individual on released results. A common formal definition is:

Here, [[MATH=M]] is a randomized mechanism, [[MATH=D]] and [[MATH=D′]] are neighboring datasets differing in one person’s record, [[MATH=S]] is any possible output set, and [[MATH=ε]] and [[MATH=δ]] are privacy parameters. Smaller [[MATH=ε]] generally provides stronger privacy but may reduce statistical utility. The definition applies under a specified neighboring-dataset model and requires careful accounting across repeated queries.

• Federated learning: trains across distributed data locations while sharing model updates rather than raw records; updates may still leak information and often require additional protections.
• Synthetic data: creates artificial records, but synthetic data can reproduce confidential patterns or fail to represent important groups.
• Privacy impact assessment: identifies necessity, proportionality, risks, safeguards, and residual risk before processing begins.

Privacy is not achieved by simply deleting names. Analysts must examine linkage, membership inference, model inversion, unintended memorization, and access pathways.

5. Security and Robustness

ML security includes conventional cybersecurity and attacks specific to data-driven systems. Threats include:

• Data poisoning: malicious or corrupted training examples alter model behavior.
• Evasion: an attacker changes an input at inference time to avoid detection.
• Model extraction: repeated queries are used to approximate a proprietary model.
• Membership inference: an attacker attempts to determine whether a person’s record was in training data.
• Prompt or instruction manipulation: relevant to systems using language models and external tools.
• Supply-chain compromise: libraries, datasets, pretrained models, or APIs introduce vulnerabilities.

A threat model should identify assets, actors, attack surfaces, assumptions, likely impacts, and controls. Controls include dataset provenance, input validation, secure development, dependency scanning, least-privilege access, rate limiting, logging, adversarial testing, backup and rollback procedures, and incident response.

Robustness also concerns ordinary changes rather than malicious attacks. Data drift is a change in input distributions; concept drift is a change in the relationship between inputs and outcomes. Monitoring should track data quality, missingness, feature distributions, prediction distributions, outcome performance, subgroup performance, latency, and override rates. Alerts require predefined actions; a dashboard without escalation authority is not effective governance.

6. Transparency, Explainability, and Contestability

Transparency means that relevant stakeholders can understand what a system does, why it is used, what data it requires, and what limitations apply. Explainability concerns understandable reasons for a particular output or model behavior. These are related but not identical.

Useful transparency artifacts include:

• plain-language notices to affected people;
• data and model documentation;
• descriptions of intended use and prohibited use;
• performance results by relevant subgroup;
• uncertainty, limitations, and known failure modes;
• information about human review and appeal procedures.

A technically plausible explanation is not automatically a legally or ethically adequate explanation. Feature importance may describe association rather than causation. Local explanations can be unstable, and simplified surrogate models may misrepresent a complex model. Explanations should therefore be validated for fidelity, stability, comprehensibility, and actionability.

Contestability gives an affected person a meaningful route to ask questions, provide additional information, request correction, appeal a decision, and obtain human reconsideration where appropriate. A human nominally placed in the loop is ineffective if the person lacks time, authority, training, or access to the evidence behind the output.

7. Accountability and Organizational Roles

Accountability requires identifiable responsibility, not merely a general statement that “the algorithm decided.” A governance structure may include:

• Business owner: defines the purpose, acceptable risk, resources, and decision rights.
• Data steward: manages provenance, quality, access, retention, and lawful use.
• Model developer: builds and documents the model and its limitations.
• Independent validator: tests assumptions, performance, fairness, security, and reproducibility.
• Privacy and legal specialists: assess applicable obligations and rights.
• Information-security team: evaluates threats and controls.
• Human decision-maker: reviews cases within defined authority.
• Internal audit or risk committee: tests governance effectiveness and evidence.

Separation between development and independent validation reduces conflicts of interest. A governance committee should approve high-risk uses, require remediation, and possess authority to pause or retire systems.

An audit trail should record model version, code and dependencies, data snapshots or hashes, training dates, parameter settings, validation results, approvals, deployment changes, input and output identifiers where lawful, human overrides, incidents, and retirement decisions. Logs must themselves be access-controlled and retention-limited. Logging everything without a purpose can increase privacy and security risk.

8. Model Validation and Documentation

Validation asks whether a model is fit for its intended use, not merely whether it performs well on a test set. A validation plan should address:

1. Conceptual soundness: Is the target a defensible representation of the business or social objective?
2. Data quality: Are data relevant, representative, sufficiently complete, and free from leakage?
3. Performance: Are suitable metrics reported on temporally and operationally realistic data?
4. Fairness: Are subgroup and intersectional effects examined?
5. Robustness: Does performance remain acceptable under plausible variation and attack?
6. Explainability: Can users understand and appropriately act on outputs?
7. Operational fit: Are latency, integration, monitoring, fallback, and human-review arrangements adequate?
8. Reproducibility: Can an independent reviewer reproduce material results?

For imbalanced classification, accuracy may be misleading. Precision, recall, specificity, area under the precision–recall curve, calibration, and cost-sensitive measures may be more relevant. Metric choice must reflect the consequences of false positives and false negatives.

A model card is a concise document containing intended use, out-of-scope use, training data, evaluation data, performance, subgroup results, limitations, ethical considerations, monitoring requirements, and contact or ownership information. A model card is not a substitute for detailed technical documentation, legal review, or an audit trail; it is an accessible communication layer.

A data sheet or dataset record should describe data origin, collection method, population, variables, labeling process, missingness, known biases, licensing, consent or legal basis, transformations, and recommended uses. Together, these artifacts make assumptions visible and support handover across teams.

9. Human Review and Decision Design

Human review is appropriate when stakes are high, uncertainty is material, cases are unusual, or legal rights are affected. It should be designed rather than asserted. Effective review specifies:

• which cases require review, such as low-confidence or adverse outcomes;
• what evidence the reviewer may examine;
• whether the reviewer can override the system;
• how reasons for overrides are recorded;
• training on automation bias and discrimination;
• service-level targets so review is timely;
• escalation and appeal routes;
• periodic evaluation of reviewer consistency and workload.

Human review can introduce its own inconsistency and bias. Therefore, organizations should compare automated recommendations, human decisions, override patterns, outcomes, and subgroup effects. For low-risk recommendations, users may simply be informed that an output is algorithmic. For high-risk decisions, meaningful human control should include authority, competence, independence, and accountability.

10. Regulatory and Legal Risk

Regulatory obligations vary by jurisdiction and sector. Relevant areas may include data protection and privacy law, anti-discrimination and employment law, consumer protection, financial regulation, sectoral safety rules, intellectual property, records management, and emerging AI-specific regulation. The correct approach is not to memorize one global rulebook but to conduct a jurisdiction- and use-case-specific assessment.

Common regulatory risk questions include:

• Is there a lawful basis and appropriate notice for processing?
• Are special categories or sensitive attributes handled lawfully?
• Does the person have rights to access, correction, objection, appeal, or review?
• Is automated decision-making restricted in this context?
• Are impact assessments, registration, documentation, or conformity evaluations required?
• Are claims about accuracy, fairness, or automation misleading?
• Are vendors contractually responsible for security, data use, and incident notification?

Organizations should maintain a regulatory register linking each obligation to a control, evidence owner, review frequency, and escalation path. Compliance is necessary but does not exhaust ethical responsibility: an action may be legal and still unreasonable, unfair, or harmful.

11. Illustrative Case: Automated Credit-Limit Recommendations

Background

A retail bank develops an ML model that recommends credit-limit increases for existing customers. The bank intends to improve customer service and reduce manual processing. The model uses transaction history, repayment behavior, account tenure, and geographic indicators. A human employee approves most recommendations, but low-risk increases are automatically issued.

Problem

After deployment, analysts observe that customers in some neighborhoods receive fewer increases. The model has good overall predictive performance, but the bank cannot clearly explain individual recommendations. An external complaint alleges that geographic variables act as proxies for ethnicity and socioeconomic status.

Analysis

The system is high risk because it influences access to credit, operates at scale, and can cause financial harm. The team examines selection rates, approval rates, repayment outcomes, false-positive and false-negative rates, calibration, and intersectional groups. It discovers that the historical label—whether a previous increase was profitable—reflects earlier staff decisions and therefore embeds historical allocation patterns. Geographic variables add predictive power but also create substantial group disparities.

Privacy review finds that transaction data were collected for account servicing, but the original notice did not clearly explain automated credit-limit recommendations. Security review identifies overly broad access to model inputs and insufficient monitoring for unusual query activity.

Approach and solution

The bank takes the following actions:

1. Suspends automatic increases while the assessment is completed.
2. Removes or constrains geographic proxies and redesigns the target around repayment capacity rather than historical profitability alone.
3. Adds subgroup and intersectional fairness thresholds to validation gates.
4. Uses a conservative human-review queue for adverse or uncertain cases.
5. Provides customers with a plain-language reason, relevant factors, and a correction and appeal process.
6. Produces a model card, data sheet, validation report, and approval record.
7. Restricts access, encrypts sensitive data, and adds audit logging.
8. Monitors performance, selection rates, complaints, overrides, and drift monthly.
9. Assigns the business owner responsibility for outcomes and requires independent validation before material changes.

Outcome

The revised model has slightly lower aggregate predictive performance but materially improves documentation, fairness analysis, customer contestability, and operational control. The bank accepts the trade-off because the original performance gain did not justify the legal and social risk. Automatic decisions remain limited to narrowly defined, low-risk cases.

Learning points

The case illustrates that responsible governance begins with purpose and labels, not only with model choice. Fairness metrics must be interpreted alongside causal history and legal context. Human review is valuable only when it is empowered and monitored. Finally, lowering automation can be a successful governance outcome rather than a technical failure.

12. Implementation Roadmap and Best Practices

A practical lifecycle contains the following gates:

1. Initiation: define purpose, affected people, alternatives to ML, and risk tier.
2. Data assessment: establish provenance, legal basis, quality, representation, minimization, and retention.
3. Design: select targets, features, metrics, thresholds, human controls, and security requirements.
4. Development: use reproducible pipelines, version control, secure environments, and peer review.
5. Validation: conduct independent testing of performance, fairness, privacy, security, explainability, and operational readiness.
6. Approval: document residual risks, owners, mitigations, and conditions of deployment.
7. Deployment: use controlled release, rollback capability, access control, notices, and training.
8. Monitoring: track drift, subgroup impacts, incidents, complaints, overrides, and changing law.
9. Reassessment: review after material changes, incidents, new data, or changes in use.
10. Retirement: revoke access, preserve required records, delete or archive data appropriately, and communicate closure.

Best practice is to involve affected stakeholders early, test adverse scenarios, avoid unnecessary sensitive data, record decisions contemporaneously, reward escalation of concerns, and treat model retirement as a normal lifecycle event. Governance should be integrated into product and project management rather than added immediately before launch.

Challenges and Limitations

Responsible ML faces persistent difficulties. Protected attributes may be unavailable or legally sensitive, yet their absence can prevent fairness testing. Labels may be incomplete or socially constructed. Metrics can conflict, and improving one group-level measure may worsen another. Privacy protection can reduce data utility. Explanations can create false confidence. Human review can be expensive, inconsistent, or biased. Vendors may provide insufficient documentation, and organizational incentives may favor rapid deployment over careful evaluation.

Governance frameworks also risk becoming ceremonial checklists. A completed form does not prove fairness or accountability. Effective governance requires independent challenge, evidence, decision rights, resources, and consequences for non-compliance. Most importantly, organizations must remain willing to redesign or abandon a system when risks cannot be reduced to an acceptable level.

Contemporary Developments

Current practice is expanding from model-level evaluation to sociotechnical system evaluation: examining users, institutions, incentives, workflows, affected communities, and downstream effects. Foundation models and generative AI add concerns about hallucination, copyrighted or confidential training data, prompt injection, tool misuse, content provenance, and rapid model change. Automated monitoring and documentation tools can improve scale, but they do not replace contextual judgment.

Organizations are increasingly adopting AI inventories, algorithmic impact assessments, red-team exercises, standardized model documentation, risk committees, and continuous control monitoring. International regulatory developments emphasize risk classification, transparency, data governance, technical documentation, human oversight, accuracy, robustness, and cybersecurity. Because requirements evolve, governance programs should be designed for change and linked to legal and policy review.

Chapter Summary

Responsible machine learning evaluates more than predictive accuracy. It requires attention to fairness, privacy, security, transparency, explainability, contestability, accountability, and regulatory obligations. Fairness is multidimensional, and metrics such as selection rates, disparate-impact ratios, true-positive rates, false-positive rates, and calibration answer different questions that may conflict. Privacy requires purpose limitation, minimization, retention controls, security, and careful assessment of de-identification; techniques such as differential privacy and federated learning reduce but do not eliminate risk.

A risk-based governance lifecycle classifies systems according to impact, autonomy, scale, and reversibility. It uses independent validation, data sheets, model cards, audit trails, human review, monitoring, incident response, and retirement procedures. Effective governance assigns named responsibilities and gives reviewers authority to pause or change a system. The credit-limit case demonstrates that responsible design may require changing labels, removing proxies, reducing automation, improving contestability, and accepting a modest reduction in aggregate performance. Ultimately, responsible ML is an ongoing sociotechnical practice grounded in evidence, proportionality, and accountability.

Key Terms

Responsible machine learning: The design, deployment, and oversight of ML systems in ways that address societal, legal, ethical, and operational risks.

Algorithmic accountability: The assignment of responsibility and answerability for system decisions, impacts, and failures.

Audit trail: A protected record of data, model, deployment, decision, approval, and change events.

Calibration: The correspondence between predicted probabilities and observed frequencies.

Concept drift: A change over time in the relationship between inputs and outcomes.

Data minimization: Collecting and using no more personal data than necessary for a defined purpose.

Data poisoning: Manipulation or corruption of training data to influence model behavior.

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

Disparate impact: A substantially unequal effect on a protected group arising from a practice.

Explainability: The extent to which a model’s behavior or output can be understood.

Fairness: A family of principles concerning justified and non-discriminatory treatment.

Human review: Deliberate assessment by an authorized and competent person of an ML-supported decision.

Model card: A structured document describing a model’s purpose, performance, limitations, and responsible-use conditions.

Model extraction: An attack that uses queries to reproduce or approximate a model.

Proxy discrimination: Discrimination produced through variables correlated with protected characteristics.

Risk classification: Categorization of an ML use according to potential harm and required controls.

Selection rate: The proportion of a group receiving a positive predicted decision.

Sociotechnical system: A system understood through the interaction of technology, people, institutions, and social context.

True-positive rate: The proportion of actual positive cases correctly identified.

Validation: Independent assessment of whether a model is fit for its intended use.

Review Questions

Short Answer Questions

1. Why is predictive accuracy insufficient as the sole criterion for responsible ML?
2. Distinguish pseudonymization from anonymization.
3. What is the purpose of a model card?
4. Define concept drift and give one example.
5. Why should human review be designed and monitored rather than merely declared?

Long Answer Questions

1. Explain how a business should evaluate fairness before deploying a high-impact classification model. Include metrics, data issues, and governance decisions.
2. Describe a complete ML governance lifecycle from initiation to retirement.
3. Discuss the relationship among transparency, explainability, and contestability.
4. Explain how privacy and security risks differ in ML systems and identify controls for each.
5. Evaluate the advantages and limitations of risk-based AI regulation and organizational governance.

Analytical / Application Questions

1. A hiring model selects 60% of applicants in Group A and 45% in Group B. Calculate the selection-rate ratio for Group B relative to Group A and explain why the result alone cannot establish discrimination.
2. A fraud model has equal accuracy across two groups but a much higher false-positive rate for one group. What harms may result, what additional metrics should be examined, and what interventions could be considered?
3. Design a governance plan for an ML system that recommends employee promotions. Specify its risk tier, validation requirements, human-review process, documentation, monitoring, and appeal mechanism.

Multiple Choice Questions with Answers

Question 1

Which statement best describes responsible machine learning?

A. Maximizing accuracy regardless of consequences

B. Using only interpretable algorithms

C. Managing technical, legal, ethical, and operational risks throughout the lifecycle

D. Allowing business users to override all model outputs

Correct Answer: C. Managing technical, legal, ethical, and operational risks throughout the lifecycle

Question 2

Which metric is defined as P(Ŷ = 1 | G = g)?

A. Selection rate

B. False-positive rate

C. Calibration error

D. Specificity

Correct Answer: A. Selection rate

Question 3

A model has a selection rate of 0.45 for Group B and 0.60 for Group A. What is the selection-rate ratio for Group B relative to Group A?

A. 0.15

B. 0.75

C. 1.33

D. 2.40

Correct Answer: B. 0.75

Question 4

Which statement about fairness criteria is correct?

A. All fairness criteria can always be satisfied simultaneously

B. Fairness requires using protected attributes as model features

C. Different criteria address different normative and statistical concerns and may conflict

D. Equal accuracy proves absence of discrimination

Correct Answer: C. Different criteria address different normative and statistical concerns and may conflict

Question 5

What does data minimization require?

A. Keeping all data indefinitely for future modeling

B. Using only data necessary for a specified purpose

C. Deleting all data after model training

D. Removing all categorical variables

Correct Answer: B. Using only data necessary for a specified purpose

Question 6

Which is an example of a data-poisoning attack?

A. Changing an input at prediction time to evade detection

B. Using repeated queries to copy a model

C. Injecting manipulated records into training data

D. Encrypting a dataset before storage

Correct Answer: C. Injecting manipulated records into training data

Question 7

What is a model card primarily intended to provide?

A. A replacement for all legal review

B. A concise description of intended use, performance, limitations, and risks

C. A database of individual customer records

D. A guarantee that the model is unbiased

Correct Answer: B. A concise description of intended use, performance, limitations, and risks

Question 8

Which condition makes human review meaningful?

A. The reviewer must accept every model recommendation

B. The reviewer has authority, competence, relevant evidence, and time to reconsider the case

C. The reviewer sees only the final score

D. The organization does not record overrides

Correct Answer: B. The reviewer has authority, competence, relevant evidence, and time to reconsider the case

Question 9

Which event is most directly associated with concept drift?

A. A change in the relationship between features and outcomes over time

B. A password expiration event

C. A model’s source code being version-controlled

D. A dataset being stored in encrypted form

Correct Answer: A. A change in the relationship between features and outcomes over time

Question 10

Why is an audit trail important?

A. It guarantees that a model is fair

B. It eliminates the need for monitoring

C. It provides evidence about decisions, changes, approvals, and incidents

D. It permits unrestricted access to personal data

Correct Answer: C. It provides evidence about decisions, changes, approvals, and incidents