5. Algorithmic Fairness and Non-Discrimination

Introduction: Why Fairness Is an Engineering and Governance Problem

Artificial intelligence systems increasingly influence access to employment, credit, education, insurance, healthcare, public benefits, policing, and digital services. These systems may classify applicants, rank candidates, recommend interventions, or allocate scarce resources. Because such decisions can affect life opportunities, accuracy alone is not an adequate criterion for evaluating an AI system. A model can be statistically accurate overall while systematically disadvantaging a particular social group.

Algorithmic fairness concerns whether the design, data, operation, and consequences of an automated system treat people justly. Non-discrimination is closely related but is also a legal and institutional obligation. It requires organizations to avoid unjustified differential treatment, disparate impacts, and the reproduction of protected-group disadvantage. Fairness is therefore not merely a technical property of a model. It is a socio-technical property shaped by historical conditions, organizational objectives, institutional rules, data practices, and human decisions surrounding the model.

The central difficulty is that fairness has several legitimate meanings. Equal selection rates, equal error rates, equal probabilities of success, and equal treatment of similarly situated individuals are not identical requirements. In many real settings they cannot all be satisfied simultaneously. Responsible practice therefore requires explicit justification of the chosen fairness objective, measurement of its consequences, and continuing governance after deployment.

Learning Objectives

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

1. Explain how historical, sampling, measurement, aggregation, and deployment choices create algorithmic bias.

2. Distinguish group fairness, individual fairness, procedural fairness, and outcome-oriented fairness.

3. Calculate and interpret demographic parity, disparate impact, equal opportunity, equalized odds, predictive parity, and individual similarity metrics.

4. Explain why competing fairness criteria may be mathematically incompatible.

5. Analyze intersectional impacts that are hidden by broad demographic categories.

6. Select suitable preprocessing, in-processing, and post-processing mitigation strategies.

7. Evaluate fairness interventions in relation to accuracy, privacy, transparency, legality, and organizational accountability.

8. Design a practical fairness assessment and monitoring process.

Bias Enters Before the Model

Algorithmic bias is a systematic pattern of error or disadvantage associated with particular individuals or groups. It may arise from the data, the target being predicted, the model, the decision threshold, or the institutional context in which predictions are used. Treating bias as a defect located only in the algorithm is misleading.

Historical bias occurs when the world represented in the training data reflects unequal institutions or past discrimination. A hiring dataset may show fewer women in senior technical roles because of earlier exclusion. Replicating that pattern can make the system appear empirically consistent while preserving an unjust arrangement.

Representation bias arises when some populations are under-sampled, absent, or represented in ways that do not reflect the deployment population. A speech-recognition system trained mainly on one accent group may perform well in laboratory testing yet fail for other speakers. Sampling bias can also occur when data are collected only from people who successfully accessed a service.

Measurement bias results when variables do not measure the underlying concept equally across groups. Criminal justice data, for example, may record arrests rather than actual offending; arrests are influenced by surveillance practices and policing patterns. A label can therefore encode institutional activity rather than the construct that a model supposedly predicts.

Label bias occurs when the outcome used as the target is incomplete, subjective, or systematically assigned. Performance ratings may reflect manager expectations as much as employee performance. Proxy discrimination occurs when apparently neutral variables, such as postcode, browsing patterns, or school attended, encode protected attributes or social disadvantage.

Aggregation bias appears when one model is applied to populations with different relationships between features and outcomes. A single risk model may be inappropriate if clinical indicators have different meanings across age groups or if language patterns vary across communities. Evaluation bias occurs when performance is assessed using a benchmark that does not represent the intended users.

Finally, deployment bias arises when a model is used in a context different from the one assumed during development. A prediction intended to support human review may become an automatic rejection rule. Feedback loops can then amplify disadvantage: people denied a loan cannot generate repayment data, and people subjected to increased surveillance generate more recorded incidents, which may be interpreted as evidence that the system was correct.

The Conceptual Landscape of Fairness

Group fairness

Group fairness evaluates whether statistical outcomes are comparable across legally or socially relevant groups. It is useful when institutions seek to detect systematic disparities, but it can conceal important variation within groups.

Demographic parity, also called statistical parity, requires equal positive decision rates:

P(Ŷ = 1 | A = a) = P(Ŷ = 1 | A = b),

where Ŷ is the model decision and A is a protected attribute. If 40 percent of Group A and 20 percent of Group B receive offers, demographic parity is not satisfied. The criterion does not ask whether the groups have equal qualifications or whether predictions are equally accurate.

The disparate-impact ratio is often written as:

DI = P(Ŷ = 1 | unprivileged group) / P(Ŷ = 1 | privileged group).

A ratio of 1 indicates equal selection rates. A ratio of 0.8 means the unprivileged group is selected at 80 percent of the privileged-group rate. Such thresholds may be used as screening indicators in organizational and legal contexts, but they are not universal proof of discrimination or fairness; context, justification, and evidence remain essential.

Equal opportunity requires equal true-positive rates:

TPR_a = P(Ŷ = 1 | Y = 1, A = a).

It asks whether genuinely qualified or eligible members of different groups are correctly selected at comparable rates. Equalized odds is stronger: it requires both equal true-positive rates and equal false-positive rates. This is relevant where both wrongful exclusion and wrongful inclusion matter.

Predictive parity requires equal positive predictive value:

PPV_a = P(Y = 1 | Ŷ = 1, A = a).

It asks whether a positive decision has the same meaning across groups. For example, among people classified as high risk, the proportion who actually experience the outcome should be comparable. Calibration is a related property: among people assigned the same predicted probability, observed outcome frequencies should be similar across groups.

Individual and procedural fairness

Individual fairness proposes that similar individuals should receive similar treatment. Formally, if a task-specific distance function d(x_i, x_j) indicates that two people are similar in relevant respects, the decision rule should not produce sharply different outcomes without a defensible reason. The challenge is that similarity is not self-defining. Choosing which attributes matter is a normative and institutional decision.

Procedural fairness examines how decisions are made rather than only their statistical results. Relevant questions include whether applicants received notice, whether data were accurate, whether explanations and appeals were available, whether decision-makers followed consistent rules, and whether affected communities had meaningful participation. Two systems with identical outcome statistics may differ substantially in procedural legitimacy.

Equality, equity, and corrective fairness

Equality generally means applying the same rule or providing the same resource. Equity recognizes that equal treatment can preserve unequal starting conditions and may require corrective measures. A fairness intervention that improves access for an underrepresented group is not necessarily unfair to others; its justification depends on the social objective, legal framework, proportionality, and evidence of disadvantage.

Measuring Fairness in a Confusion-Matrix Framework

For a binary classifier, the confusion matrix contains true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). From it we calculate:

• True-positive rate or recall: TP/(TP+FN).
• False-positive rate: FP/(FP+TN).
• True-negative rate or specificity: TN/(TN+FP).
• Positive predictive value or precision: TP/(TP+FP).
• Selection rate: (TP+FP)/N.

Consider the following illustrative results for a recruitment screening system:

| Group | TP | FP | TN | FN | Total |

|---|---:|---:|---:|---:|---:|

| Group A | 36 | 12 | 48 | 4 | 100 |

| Group B | 24 | 16 | 44 | 16 | 100 |

For Group A, the selection rate is 48/100 = 0.48, TPR is 36/40 = 0.90, FPR is 12/60 = 0.20, and PPV is 36/48 = 0.75. For Group B, the selection rate is 40/100 = 0.40, TPR is 24/40 = 0.60, FPR is 16/60 ≈ 0.267, and PPV is 24/40 = 0.60.

The disparate-impact ratio for B relative to A is 0.40/0.48 ≈ 0.83. The system therefore shows a selection-rate disparity, lower opportunity for qualified members of B, a higher false-positive rate for B, and lower predictive precision for B. No single metric summarizes the entire situation. The appropriate response depends on the decision’s purpose and the harms attached to each type of error.

Fairness metrics should be accompanied by uncertainty estimates, subgroup sample sizes, threshold information, and practical harm analysis. Small groups can produce unstable rates. A difference may be statistically detectable but practically minor, or practically serious despite limited statistical power. Metrics should be computed on relevant time periods and monitored after deployment because population composition and model performance change.

Why Fairness Criteria Conflict

Suppose two groups have different base rates, meaning different proportions of positive outcomes in the population. If a model is calibrated and has equal predictive value across groups, equalizing error rates may require group-specific thresholds. Conversely, imposing equal selection rates can alter calibration and predictive value. Under common conditions involving unequal base rates and a non-perfect predictor, demographic parity, equalized odds, and predictive parity cannot all hold simultaneously.

This is not a technical failure that can always be solved by finding a better classifier. It is a decision about which errors and distributions of opportunity are ethically acceptable. In lending, an institution may prioritize equal access and therefore examine selection rates and false-negative disparities. In medical screening, missing a true case may be more harmful than generating additional referrals, making equal opportunity especially important. In public-sector decisions, due process, contestability, and protection against arbitrary classification may be central.

Fairness metrics also differ in their causal assumptions. A disparity may result from discrimination, unequal access to preparation, legitimate risk differences, measurement error, or a mixture of these. Statistical parity can reveal an important warning signal but does not by itself establish the cause. Causal fairness approaches ask whether protected attributes or their descendants have an impermissible causal influence, while recognizing that protected attributes may sometimes be needed to detect and correct discrimination.

Intersectionality: When Categories Hide the Harm

Intersectionality examines how multiple identities and social positions combine. A model may show acceptable outcomes for women overall and for racial minority applicants overall while severely disadvantaging women from a particular racial minority. Aggregated analysis can conceal this pattern because the group averages combine people with very different experiences.

An intersectional audit should examine combinations such as race and gender, disability and age, language and immigration status, or region and socioeconomic position, subject to privacy and statistical reliability constraints. Analysts should avoid indiscriminate subgroup slicing that creates unstable estimates or exposes sensitive information. Where sample sizes are small, confidence intervals, qualitative investigation, pooled hierarchical models, privacy-preserving analysis, and community consultation can complement simple rate comparisons.

Intersectionality also has a design implication: affected communities should participate in defining harms and relevant outcomes. A metric selected by a central analytics team may fail to represent the lived consequences of a system’s errors.

A Fairness Assessment Workflow

A credible fairness assessment begins by defining the decision, the affected populations, the decision-maker’s authority, and the potential harms. The team should identify protected and vulnerable attributes, including attributes that may not be used as model inputs but are necessary for auditing. Data collection must have a lawful and ethical basis, with access controls and safeguards against re-identification.

The next step is to examine the target variable and its provenance. Analysts should ask who created the label, what incentives shaped it, what outcomes are unobserved, and whether it is equally meaningful across groups. Exploratory analysis should compare missingness, representation, feature distributions, selection rates, error rates, and intersectional outcomes.

Model evaluation should use a holdout set that reflects intended deployment. It should report overall performance and subgroup performance, threshold sensitivity, confidence intervals, calibration, and operational consequences. Qualitative review is valuable: affected users may identify harms that numerical metrics overlook.

After deployment, governance should include monitoring for drift, complaints, appeals, override patterns, subgroup error changes, and feedback loops. A model should have an accountable owner, documented change controls, escalation procedures, and a defined suspension or withdrawal mechanism. Fairness is not a one-time certification.

Mitigation Across the Machine-Learning Lifecycle

Preprocessing

Preprocessing changes data before model training. Re-sampling can increase representation of an underrepresented group, though naïve duplication may cause overfitting. Reweighting assigns different importance to examples so that the learner pays greater attention to underrepresented or disadvantaged cases. Label review can correct inconsistent or biased annotations. Feature transformation can reduce reliance on obvious proxies, although removing protected attributes alone rarely solves the problem because correlated variables remain.

More advanced methods learn a transformed representation that preserves predictive information while reducing information associated with a protected attribute. Synthetic data may help explore rare cases, but it can also reproduce existing bias or create unrealistic records. Data augmentation and targeted collection are often preferable when feasible. Preprocessing does not remove the need for post-deployment monitoring.

In-processing

In-processing incorporates fairness into model training. A constrained optimization approach may maximize predictive performance subject to limits on disparities in selection rates or error rates. A regularized objective can penalize measured unfairness:

Loss = predictive loss + λ(fairness penalty),

where λ represents the organization’s chosen trade-off. Adversarial debiasing trains a predictor while an adversary attempts to infer protected attributes from its representation; the predictor is optimized to retain task information while reducing that inferability. These methods require careful specification because suppressing information about a protected attribute can sometimes reduce the ability to identify and correct group disadvantage.

In-processing techniques can be powerful but may be difficult to explain, reproduce, or apply when organizations use proprietary models. Constraints may also reduce performance or shift errors to groups not included in the audit.

Post-processing

Post-processing modifies model scores, thresholds, or final decisions after training. Group-specific thresholds can equalize TPR, FPR, or another selected criterion. Reject-option classification may defer borderline cases for human review. A calibrated model may support policy-based decision rules rather than a single automatic cutoff.

Post-processing is practical when the underlying model cannot be retrained, but it can complicate transparency and legal analysis. Affected individuals may receive different thresholds, and the organization must be able to explain why this is necessary and proportionate. Post-processing also cannot repair a fundamentally invalid target or a deployment process that produces coercion or exclusion.

A sound mitigation strategy compares alternatives through a multi-objective evaluation: predictive utility, fairness metrics, subgroup and intersectional harms, transparency, privacy, operational feasibility, legal defensibility, and reversibility. “Fairness through unawareness”—simply deleting protected attributes—is generally insufficient.

Applied Case: FairHire Screening System

Background

FairHire, an illustrative technology company, supplies a résumé-screening tool to large employers. The tool produces a score predicting whether an applicant will pass a structured technical assessment. It was trained on five years of historical applications and outcomes from the company’s existing recruitment process.

Situation and problem

During an audit, the employer discovered that applicants identified as women were selected for assessment less often than men. The vendor argued that gender was not an input. Further analysis found that the model relied heavily on uninterrupted employment history, participation in particular professional networks, and names of institutions. These variables were correlated with gender and socioeconomic status. The historical process itself had produced fewer opportunities for women to reach later recruitment stages.

Analysis

Among 1,000 men, 500 were selected; among 1,000 women, 350 were selected. The selection-rate ratio for women was 0.70. Among applicants later shown by the structured assessment to be qualified, the model selected 80 percent of men but only 60 percent of women. Thus the disparity involved both overall selection and equal opportunity. A further intersectional analysis showed that women from a minority ethnic group had a 45 percent selection rate, a pattern hidden by the aggregate female category.

The audit team also questioned the label. “Successful applicant” reflected advancement through the old process, not job performance after hiring. Consequently, the target encoded historical opportunity and managerial judgments. Removing gender would not address these problems.

Strategy and solution

FairHire and the employer first replaced the historical advancement label with a validated combination of job-relevant assessment results and post-hire performance indicators, after reviewing privacy and employment-law requirements. They conducted structured label review and removed or transformed features lacking a defensible job-related rationale. Reweighting and targeted data collection improved representation. The team then trained a constrained model with an equal-opportunity objective and tested threshold alternatives.

The organization chose a policy that limited disparities in true-positive rates, introduced human review for borderline cases, and required recruiters to document overrides. It retained protected-attribute data in a controlled audit environment rather than using those attributes as ordinary predictors. Applicants received notice that automated screening was used and were given a channel to request correction or review.

Outcome

The revised system reduced the gap in selection rates and substantially narrowed the true-positive-rate disparity. Overall accuracy declined slightly, but the employer judged the change acceptable because the previous accuracy measure rewarded replication of historical decisions. Monitoring was scheduled quarterly, with additional review of intersectional groups and complaints.

Learning points

The case demonstrates that fairness analysis must examine labels, proxies, process design, and deployment—not merely model features. It also shows why group-level metrics should be combined with intersectional analysis, procedural safeguards, and governance. A technically effective mitigation can still be inadequate if applicants cannot challenge inaccurate data or if the organization lacks accountability for human overrides.

Organizational and Governance Responsibilities

Fairness work requires collaboration among data scientists, domain experts, legal and compliance teams, human-resource or public-service professionals, affected communities, and senior decision-makers. Leadership must define risk tolerance and provide resources for auditing, documentation, accessibility, and remediation. Technical teams should maintain data sheets, model cards, metric definitions, threshold records, subgroup evaluations, and known limitations.

Organizations should distinguish assistance from delegation. A model that recommends candidates is not the same as a system that automatically rejects them. Human involvement is meaningful only when reviewers have authority, time, training, and information to disagree with the system. Automation bias can cause humans to rubber-stamp model outputs.

Privacy creates a difficult tension. Protected attributes may be sensitive, yet without them organizations cannot measure disparate impact. Collection should therefore be minimized, purpose-limited, access-controlled, and separated from operational decision-making where appropriate. Fairness claims should not be made using unverified assumptions about identity or inferred demographic labels.

Legal standards differ by jurisdiction and sector. Organizations must examine applicable anti-discrimination law, employment and lending rules, accessibility duties, data-protection requirements, procurement conditions, and requirements for explanation or appeal. A fairness metric is not a substitute for legal advice, impact assessment, or institutional judgment.

Limitations and Common Errors

A common mistake is to report only overall accuracy. Another is to select a metric after inspecting results, choosing whichever produces the most favorable narrative. Teams may also confuse correlation with causation, assume that removing protected attributes removes discrimination, ignore small or intersectional groups, or treat fairness as a one-time model adjustment.

Metrics can be gamed. Equal selection rates may be achieved by lowering standards or changing labels without improving substantive opportunity. Equal error rates may be inappropriate when errors have different costs. A model can satisfy a statistical criterion while causing procedural harm through opaque denials, inaccessible appeals, or excessive data collection.

Fairness interventions can create trade-offs and unintended effects. Group-specific thresholds may stigmatize groups or produce legal complications. Reweighting may reduce performance for rare cases. Debiasing representations may remove information useful for auditing. Human review may introduce inconsistent discretion. Every intervention should therefore be tested for distributional effects, robustness, explainability, and reversibility.

Emerging Directions

Current work increasingly combines fairness metrics with causal analysis, distributionally robust learning, uncertainty quantification, and participatory design. Researchers are developing methods for fairness under dataset shift, federated settings, and privacy constraints. There is growing attention to generative AI, where bias can appear in training corpora, generated content, ranking, refusals, and evaluation benchmarks.

Another important direction is continuous algorithmic impact assessment. Rather than treating a model as an isolated artifact, organizations evaluate the entire decision system, including objectives, institutional incentives, human roles, affected communities, and environmental or social externalities. Documentation standards, independent audits, algorithm registers, procurement requirements, and incident reporting are becoming important governance mechanisms. These developments reinforce a central principle: fairness is an ongoing public and organizational responsibility, not a numerical property that a model permanently possesses.

Best-Practice Principles

A professional fairness program should:

• define the decision, harm, population, and legal context before selecting metrics;
• examine data provenance, labels, missingness, proxies, and historical access;
• report multiple metrics with uncertainty and intersectional analysis;
• involve domain experts and affected communities in defining relevant harms;
• compare preprocessing, in-processing, and post-processing interventions;
• document trade-offs rather than claiming that one metric proves fairness;
• provide notice, explanation, correction, appeal, and human review where appropriate;
• monitor drift, overrides, complaints, and subgroup outcomes after deployment;
• protect audit data and use sensitive attributes only for justified purposes; and
• assign clear accountability, including authority to pause or withdraw the system.

Chapter Summary

Algorithmic fairness addresses whether AI-supported decisions distribute opportunities, errors, and burdens justly. Bias may enter through historical institutions, unrepresentative samples, flawed labels, proxy variables, aggregation, evaluation, and deployment. Group fairness metrics include demographic parity, disparate impact, equal opportunity, equalized odds, predictive parity, and calibration. Individual fairness focuses on similar treatment for similar people, while procedural fairness emphasizes notice, accuracy, explanation, participation, and appeal.

No single fairness definition is universally correct, and several criteria conflict when group base rates differ. Fairness assessment therefore requires an explicit decision context, multiple metrics, intersectional analysis, and attention to causal and procedural factors. Preprocessing changes data, in-processing embeds constraints or penalties in training, and post-processing changes thresholds or decisions. Each creates trade-offs and must be governed throughout the model lifecycle. Effective practice combines technical measurement with privacy protection, organizational accountability, community participation, legal review, and continuous monitoring.

Key Terms

Algorithmic bias: A systematic pattern of disadvantage or error associated with a group or individual in an automated decision system.

Demographic parity: Equal positive decision rates across specified groups.

Disparate impact: An adverse difference in outcomes associated with group membership, often examined through selection-rate ratios.

Equal opportunity: Equality of true-positive rates across groups.

Equalized odds: Equality of both true-positive and false-positive rates across groups.

Predictive parity: Equality of positive predictive value across groups.

Calibration: Consistency between predicted probabilities and observed outcome frequencies across groups.

Individual fairness: The principle that similar individuals should receive similar treatment in relevant respects.

Procedural fairness: Fairness of the decision-making process, including notice, consistency, explanation, and appeal.

Historical bias: Bias inherited from unequal social or institutional conditions represented in data.

Representation bias: Bias caused by missing or unrepresentative population examples.

Measurement bias: Unequal validity of a feature or label as a measure of the intended construct.

Proxy discrimination: Disadvantage caused by variables correlated with protected attributes.

Aggregation bias: Error arising when one model combines populations with different feature-outcome relationships.

Intersectionality: Analysis of overlapping identities and social positions that may produce distinctive disadvantage.

Preprocessing: Fairness mitigation applied to data before model training.

In-processing: Fairness mitigation incorporated into the model-learning objective.

Post-processing: Fairness mitigation applied to scores, thresholds, or decisions after training.

Feedback loop: A process in which system outputs alter future data and thereby reinforce earlier patterns.

Algorithmic impact assessment: A structured evaluation of an AI system’s risks, effects, governance, and accountability.

Review Questions

Short Answer Questions

1. What is the difference between historical bias and measurement bias?

2. Define demographic parity and disparate impact ratio.

3. What does equal opportunity measure?

4. Why can deleting a protected attribute fail to eliminate discrimination?

5. What is meant by intersectional fairness analysis?

Long Answer Questions

1. Explain the major points in an AI lifecycle at which bias can enter, using suitable examples.

2. Compare demographic parity, equalized odds, predictive parity, and individual fairness.

3. Explain why fairness criteria may be incompatible when base rates differ.

4. Describe preprocessing, in-processing, and post-processing mitigation methods, including their trade-offs.

5. Design a governance process for monitoring fairness in an automated recruitment system.

Analytical / Application Questions

1. A model selects 60 percent of Group A and 45 percent of Group B. Calculate the selection-rate ratio for B and explain what additional information is needed before judging the system.

2. In a medical screening system, Group X has a lower true-positive rate but the same false-positive rate as Group Y. Which fairness criterion is violated, and what harm might result?

3. A system appears fair for men and women separately but performs poorly for women from a minority ethnic group. Explain how an audit should investigate and respond to this finding.

22. Multiple Choice Questions with Answers

Question 1

A recruitment model selects 48% of Group A and 40% of Group B. What is the disparate-impact ratio for Group B relative to Group A?

A. 0.48

B. 0.83

C. 1.20

D. 8.00

Correct Answer: B. 0.83

Question 2

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

A. Demographic parity

B. Predictive parity

C. Equal opportunity

D. Calibration

Correct Answer: C. Equal opportunity

Question 3

A model has equal true-positive rates and equal false-positive rates for two groups. Which criterion does this satisfy?

A. Equalized odds

B. Individual fairness

C. Demographic parity

D. Procedural fairness

Correct Answer: A. Equalized odds

Question 4

Why is removing gender from a model often insufficient to prevent discrimination?

A. Models cannot operate without gender

B. Other variables may act as proxies for gender

C. Removing gender always lowers accuracy to zero

D. Fairness metrics require gender as an input

Correct Answer: B. Other variables may act as proxies for gender

Question 5

Which example best illustrates measurement bias?

A. A group is absent from the training sample

B. A model is deployed in a new population

C. Arrest records are used as a direct measure of offending

D. One model is used for two populations with different relationships

Correct Answer: C. Arrest records are used as a direct measure of offending

Question 6

What is the main purpose of intersectional analysis?

A. To eliminate all subgroup analysis

B. To combine all protected attributes into one score

C. To identify harms affecting overlapping identity groups

D. To guarantee demographic parity automatically

Correct Answer: C. To identify harms affecting overlapping identity groups

Question 7

Which is an in-processing mitigation technique?

A. Reweighting training examples before learning

B. Changing decision thresholds after training

C. Adding a fairness penalty to the training objective

D. Providing applicants with an appeal process

Correct Answer: C. Adding a fairness penalty to the training objective

Question 8

A model is calibrated across groups, but groups have unequal base rates. Which conclusion is most appropriate?

A. All other fairness criteria must also hold

B. Some fairness criteria may be incompatible

C. The model is necessarily discriminatory

D. Calibration proves individual fairness

Correct Answer: B. Some fairness criteria may be incompatible

Question 9

Which practice best represents meaningful human oversight?

A. A reviewer automatically approves every model output

B. A reviewer has authority, time, training, and evidence to challenge outputs

C. A human is listed as responsible but cannot alter decisions

D. The system hides its uncertainty from reviewers

Correct Answer: B. A reviewer has authority, time, training, and evidence to challenge outputs

Question 10

Why should protected attributes sometimes be collected in a controlled audit environment?

A. To use them secretly in every prediction

B. To enable measurement of disparate outcomes and mitigation

C. To replace all other model features

D. To avoid privacy safeguards

Correct Answer: B. To enable measurement of disparate outcomes and mitigation