AI, Digital Transformation, Ethics and the Future of HR Analytics

Introduction

Human resource analytics is entering a period in which artificial intelligence (AI), cloud platforms, automation, workforce data, and digital collaboration tools are becoming integrated into everyday people management. Recruitment systems rank applicants, learning platforms recommend development content, workforce-planning tools forecast capability requirements, and generative AI drafts job descriptions, interview guides, employee communications, and analytical narratives.

These developments create substantial opportunities. HR functions can process larger datasets, identify emerging skills, personalise employee development, reduce administrative work, and provide faster evidence for strategic decisions. However, greater analytical capability also creates greater responsibility. A prediction can influence access to employment, promotion, pay, training, scheduling, or termination. If the underlying data are incomplete or historically biased, an apparently objective system may reproduce or amplify inequality.

The central issue is therefore not whether HR should use AI, but under what conditions AI and digital transformation improve human decision-making without undermining dignity, autonomy, privacy, fairness, security, or accountability. This chapter evaluates the principal technologies and establishes a governance approach for responsible HR analytics.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the role of generative AI, people analytics platforms, automation, skills intelligence, and digital transformation in HR.
2. Distinguish descriptive, predictive, prescriptive, and generative uses of AI in people management.
3. Evaluate the benefits, limitations, and risks of AI-enabled HR decisions.
4. Design controls for privacy, consent, security, fairness, transparency, human oversight, and regulatory compliance.
5. Apply a responsible-AI governance framework to an HR analytics implementation.
6. Critically assess emerging trends and the future professional role of HR analysts.

1. From HR Information Systems to Intelligent People Operations

The evolution of HR technology can be understood in four broad stages:

Digital transformation is more than replacing paper with software. It involves redesigning processes, roles, decision rights, data architectures, and employee experiences around digital capabilities. A transformation succeeds when technology, operating model, skills, governance, and organisational culture are aligned.

A useful distinction is between digitisation, digitalisation, and digital transformation. Digitisation converts analogue information into digital form. Digitalisation uses digital tools to improve an existing process. Digital transformation changes how value is created and how work and decisions are organised. Installing an applicant tracking system is digitalisation; redesigning recruitment around skills-based assessment, structured selection, predictive workforce planning, and candidate self-service is transformation.

The transformation of HR should not be measured only by the number of tools deployed. Relevant outcomes include decision quality, employee experience, inclusion, time saved, data quality, strategic contribution, and the distribution of benefits and risks across stakeholder groups.

2. The Contemporary AI and Analytics Landscape

2.1 Types of analytical intelligence

HR analytics commonly includes four forms of capability:

• Descriptive analytics explains what has happened, such as turnover by department.
• Diagnostic analytics investigates why a pattern may have occurred, such as links between workload, management practices, and exits.
• Predictive analytics estimates what may happen, such as the probability that an employee will leave.
• Prescriptive analytics recommends possible actions, such as targeted career conversations or workload interventions.

These categories are not equivalent to certainty. A prediction is an estimate based on data and assumptions, not a fact about an individual. Prescriptive outputs remain recommendations and require judgement concerning context, feasibility, and values.

2.2 Generative AI

Generative AI produces new text, code, images, summaries, recommendations, or other content from prompts and data. In HR it may assist with:

• drafting and adapting job descriptions;
• summarising employee-survey comments;
• creating learning materials and practice scenarios;
• translating communications;
• generating interview questions aligned with job criteria;
• answering policy questions through an HR assistant; and
• helping analysts write queries, documentation, and first-draft reports.

Generative AI is probabilistic. It can produce fluent but inaccurate statements, omit relevant context, reveal confidential information, or reproduce stereotypes present in training data. The phenomenon commonly called a hallucination is especially dangerous in HR because an invented policy interpretation or fabricated candidate attribute may affect rights and opportunities.

Generative AI should therefore be treated as a copilot rather than an autonomous decision-maker. Its output requires fact-checking, source verification, appropriate disclosure, and human accountability. Prompts and outputs may also constitute personal or confidential data; employees should not paste sensitive information into unapproved public tools.

2.3 People analytics platforms

A modern people analytics platform may integrate HR records, recruitment, learning, performance, compensation, engagement, time, scheduling, collaboration, and skills data. Its value depends on more than integration. The organisation must establish data ownership, common definitions, lineage, access controls, retention rules, and processes for correcting inaccurate records.

A dashboard can create an illusion of precision. A turnover rate calculated from inconsistent employee populations, changing definitions, or missing contingent workers may be visually attractive but analytically weak. Platform adoption should therefore follow data governance, not substitute for it.

2.4 Automation and intelligent workflows

Automation is appropriate where rules are stable, inputs are reliable, and consequences are limited or readily reversible. Examples include routing a leave request, reminding a manager about an overdue review, validating mandatory fields, or scheduling standard communications.

Automation becomes high risk when it makes or materially influences decisions about hiring, promotion, dismissal, pay, performance, or access to opportunity. In such settings, the organisation should provide meaningful human review, an explanation suitable for the affected person, an appeal route, and monitoring for disparate impact.

2.5 Skills intelligence

Skills intelligence uses structured taxonomies, inferred capabilities, work history, learning activity, qualifications, assessments, and labour-market information to represent organisational capability. It can support internal mobility, workforce planning, reskilling, project staffing, and more inclusive skills-based recruitment.

However, inferred skills are not necessarily possessed skills. A system may infer competence from a job title, course completion, or text in a profile without verifying practical performance. Skills taxonomies can also reflect the assumptions of dominant occupations and become obsolete as technology changes. Employees should be able to review, correct, and, where appropriate, challenge their skills profiles.

3. A Socio-Technical View of AI in HR

AI systems are socio-technical systems: outcomes arise from the interaction of models, data, users, processes, institutions, and affected people. A technically accurate model can still produce an unacceptable decision if the objective is inappropriate or the process denies due process.

The following chain is useful for analysis:

Purpose → Data → Model → Decision → Impact → Review

At the purpose stage, HR must ask what decision is being supported and whether automation is necessary. At the data stage, it must examine relevance, quality, representativeness, provenance, and legality. At the model stage, it must assess performance, explainability, robustness, and limitations. At the decision stage, it must define who may act on the output. At the impact stage, it must monitor consequences for different groups. At the review stage, it must provide correction, appeal, audit, and system improvement.

This chain discourages a narrow focus on algorithmic accuracy. A model may rank candidates accurately according to historical hiring outcomes while learning historical discrimination. Prediction of a sensitive outcome may also be ethically unacceptable even if statistically strong.

Human-in-the-loop, on-the-loop, and in-command

• Human-in-the-loop: a person reviews and approves an individual decision before action.
• Human-on-the-loop: the system operates automatically while a person monitors performance and intervenes when necessary.
• Human-in-command: accountable leaders govern the system’s purpose, boundaries, deployment, and withdrawal.

For high-impact employment decisions, human-in-the-loop review is generally stronger than passive monitoring. Yet a nominal human reviewer is insufficient if workload, interface design, or organisational pressure causes automation bias—the tendency to accept a machine output without adequate scrutiny.

4. Responsible AI Principles for HR

4.1 Purpose limitation and proportionality

Data and AI should be used for specified, legitimate purposes. An organisation should not collect extensive behavioural data merely because it might become useful later. The intrusiveness of data collection should be proportionate to the value and necessity of the decision. Monitoring keystrokes or inferred emotions, for example, requires exceptional justification and may be inappropriate even where technically possible.

4.2 Privacy and data minimisation

Privacy includes control over personal information, freedom from unjustified surveillance, and respect for contextual expectations. HR data may include identity, health, compensation, family, biometric, behavioural, and inferred information. Inferences can be sensitive even when the original variables appear ordinary.

Data minimisation requires collecting and retaining only what is necessary. Practical controls include purpose registers, retention schedules, role-based access, encryption, pseudonymisation where feasible, secure deletion, and restrictions on secondary use. Aggregation does not guarantee anonymity: small groups may be re-identified by combining datasets.

4.3 Consent and lawful basis

Consent is meaningful only when it is informed, specific, freely given, unambiguous, and capable of withdrawal without unfair disadvantage. In employment, power asymmetry may make consent difficult to regard as genuinely free. Consent is therefore not a universal solution to workplace surveillance or analytics.

Depending on jurisdiction, processing may rely on other lawful bases, such as legal obligation, contractual necessity, legitimate interests, or public interest. These bases have different requirements and do not remove duties of transparency, proportionality, security, and fairness. Organisations must obtain jurisdiction-specific legal advice rather than treating a consent notice as complete compliance.

4.4 Security and resilience

HR systems are attractive targets because they hold valuable identity and financial data. Controls should include strong authentication, least-privilege access, segregation of duties, encryption in transit and at rest, logging, vulnerability management, secure development, vendor assessment, incident response, backups, and continuity planning.

AI introduces additional threats: prompt injection, data poisoning, model extraction, insecure plug-ins, malicious uploads, and leakage through logs or model outputs. A vendor contract should address data use for training, subcontractors, breach notification, audit rights, deletion, service continuity, and portability.

4.5 Fairness and discrimination prevention

Fairness is multidimensional. Possible concerns include equal treatment, equal opportunity, equal outcomes, error-rate parity, accessibility, and procedural fairness. These principles can conflict; no single fairness metric resolves every normative question.

Organisations should test for disparate impact across legally protected and otherwise vulnerable groups, while recognising that collecting demographic information for auditing itself requires safeguards. Analysis should examine the full pipeline: job design, sourcing, data labels, feature selection, model thresholds, human review, and outcome patterns.

A common statistical measure is a selection-rate ratio, sometimes called adverse-impact or disparate-impact ratio:

Here, S_g is the selection rate for a focal group and S_r is the selection rate for a reference group. A selection rate is selected applicants divided by eligible applicants. The ratio is dimensionless. The measure is a screening indicator, not a universal legal test; legal standards vary by jurisdiction, sample sizes matter, and legitimate job-related explanations must be investigated. For example, if 30 of 100 applicants in one group are selected and 50 of 100 in a reference group are selected, R = 0.30 ÷ 0.50 = 0.60. This indicates a substantial disparity requiring investigation, not automatic proof of unlawful discrimination.

4.6 Transparency and explainability

Transparency operates at several levels:

1. Notice: people know that AI or analytics is being used.
2. Purpose: they understand what decision or service it supports.
3. Process: authorised stakeholders understand relevant inputs, limitations, and review points.
4. Outcome explanation: an affected person receives a useful, comprehensible reason for a decision.

Technical interpretability is not the same as a meaningful explanation. Listing hundreds of features does not explain a decision to an employee. A useful explanation identifies the decision criteria, relevant information, uncertainty, human role, and route for correction or appeal.

4.7 Human oversight, contestability, and accountability

Every system should have a named business owner, technical owner, privacy or legal adviser, and escalation route. Employees and candidates should be able to request correction of inaccurate data, challenge consequential decisions, and obtain human review. Human oversight must have authority to override the system and must not be merely ceremonial.

Accountability cannot be delegated to a vendor. The organisation deploying a system remains responsible for its use, even where the model is purchased as software-as-a-service.

5. Regulatory and Standards Landscape

The applicable rules depend on jurisdiction, sector, data type, and use case. HR leaders should distinguish law from voluntary standards and should conduct a documented legal assessment before deployment.

• Data-protection law: Regimes such as the EU General Data Protection Regulation establish principles concerning lawfulness, fairness, transparency, purpose limitation, data minimisation, accuracy, storage limitation, security, and accountability. They may also regulate automated decision-making, special-category data, impact assessments, and individual rights.
• Employment and equality law: Anti-discrimination, equal-pay, disability, labour, collective-bargaining, and occupational-safety rules may apply to AI-assisted employment decisions. A system can be unlawful even when no programmer intended discrimination.
• AI-specific regulation: The European Union AI Act adopts a risk-based approach and treats many employment, worker-management, and access-to-work uses as high risk, subject to obligations that may include risk management, data governance, documentation, logging, human oversight, accuracy, robustness, and cybersecurity. Prohibited practices and obligations depend on the precise use case and legal role of the organisation.
• Standards and guidance: The NIST AI Risk Management Framework, ISO/IEC 42001, ISO/IEC 23894, and related standards provide useful governance structures, but certification or alignment does not automatically establish legal compliance.

A compliance register should identify each system, purpose, data category, legal basis, jurisdiction, risk classification, vendor, retention period, assessment status, owner, and review date. Data-protection impact assessments and algorithmic impact assessments are valuable before deployment, particularly for high-impact processing.

6. A Responsible AI Governance Lifecycle

A practical lifecycle consists of seven linked stages.

Stage 1: Define and justify

Describe the business problem, affected stakeholders, intended benefit, prohibited uses, and decision boundary. Establish a baseline process so that the organisation can compare whether AI genuinely improves outcomes.

Stage 2: Screen risk

Classify the use case according to impact, sensitivity, scale, reversibility, vulnerable populations, and degree of automation. High-risk examples include candidate ranking, promotion recommendations, termination risk scores, emotion inference, and productivity surveillance.

Stage 3: Assess data and alternatives

Document data provenance, quality, representativeness, missingness, labels, proxies, retention, and access. Consider whether a less intrusive method, such as structured human assessment, would achieve the objective.

Stage 4: Validate before deployment

Conduct technical, legal, ethical, security, and user testing. Evaluate accuracy and subgroup performance. Test robustness to changed labour markets and altered work practices. Use a pilot with predefined success and stop criteria.

Stage 5: Deploy with controls

Provide training, access restrictions, human review, explanation notices, appeal routes, incident procedures, and change-management support. Prohibit unapproved tools and unreviewed generative-AI output in official employment decisions.

Stage 6: Monitor continuously

Track data drift, model drift, performance, subgroup outcomes, complaints, overrides, security events, and unintended work intensification. Monitoring should include qualitative feedback from employees and managers, not only technical metrics.

Stage 7: Review, renew, or retire

Reassess the system after material changes in data, law, workforce, vendor, or purpose. Retire tools that no longer provide net value or cannot be governed adequately. Decommissioning must include data deletion or return and preservation of required records.

7. Implementation Strategy for HR Leaders

Implementation is an organisational change programme rather than an IT installation. A mature strategy includes:

1. Executive sponsorship and accountability: establish a cross-functional AI governance committee involving HR, legal, privacy, information security, IT, employee representatives, and affected business units.
2. Use-case portfolio: rank proposed applications by value, feasibility, and risk. Begin with low-risk, reversible applications such as policy search or administrative workflow support before high-impact decision systems.
3. Data architecture: establish common definitions, master data, lineage, metadata, identity management, and quality controls.
4. Capability development: train HR professionals in data literacy, prompt competence, bias awareness, privacy, model limitations, and responsible challenge.
5. Change and communication: explain why a tool is being used, what it cannot do, how employees are protected, and how feedback will shape the system.
6. Procurement discipline: require documentation, testing access, security evidence, accessibility, audit rights, service levels, and clear allocation of responsibilities.
7. Outcome evaluation: assess not only efficiency but also fairness, employee trust, decision quality, inclusion, and unintended consequences.

Evaluating a proposed use case

A decision committee can ask:

• Is the purpose legitimate, necessary, and sufficiently specific?
• What is the least intrusive method that could achieve it?
• Who may be harmed by an error or inference?
• Can the data and model be explained and challenged?
• Does a human have the competence and authority to override the output?
• What evidence will show benefit, fairness, and safety?
• What conditions require suspension or withdrawal?

8. Illustrative Case: Skills-Based Internal Mobility at Northstar Manufacturing

Background. Northstar Manufacturing, a fictional multinational producer, faces shortages in digital maintenance, industrial data analysis, and low-carbon production skills. It proposes an AI-enabled internal marketplace that infers employee skills from job histories, learning records, project documents, and manager profiles, then recommends vacancies and training.

Problem. The initial pilot produces fewer recommendations for part-time employees, workers with older job titles, and employees whose profiles are written in a non-dominant language. Managers also begin using the recommendations as informal promotion rankings.

Analysis. The system’s objective is internal mobility, but its data reflect unequal access to training, inconsistent profile writing, historical occupational segregation, and language differences. An inferred absence of evidence is being treated as evidence of absence. The repurposing of recommendations as promotion rankings creates a higher-impact use than the approved design.

Approach and solution. Northstar takes five actions. First, it separates skills discovery from promotion decisions and prohibits managers from using the score as a performance rating. Second, it allows employees to inspect, correct, and add evidence to their skills profiles. Third, it supplements inferred skills with structured assessments, manager validation, and employee self-reporting. Fourth, it audits recommendation rates and quality across relevant groups, including accessibility and language effects. Fifth, it introduces human career advisers, a documented appeal process, and quarterly governance reviews.

The organisation also minimises data by excluding private collaboration content that is not necessary for skills matching. It communicates the purpose and limits of inference, obtains appropriate permissions or establishes another lawful basis, and contracts with the vendor to prevent unauthorised secondary use.

Outcome. In this illustrative case, recommendation coverage improves for previously underrepresented groups, while managers receive clearer guidance that recommendations are developmental rather than determinative. Some efficiency gains are reduced because additional validation is required, but employee trust and procedural fairness improve. Northstar decides that the system is appropriate for discovery and learning recommendations but not for autonomous promotion decisions.

Learning points. The case demonstrates that fairness is affected by process design, not only model code. It also shows why purpose limitation, employee participation, contestability, and risk-based boundaries are essential to responsible skills intelligence.

9. Challenges and Limitations

Data and measurement limitations

HR data are often fragmented, historically produced, and shaped by managerial discretion. Job titles may not represent actual work; performance ratings may contain bias; engagement surveys may suffer from non-response; and turnover labels may obscure involuntary exits or labour-market conditions. Better algorithms cannot fully correct invalid constructs.

Proxy discrimination

A model may exclude protected attributes yet use proxies such as postcode, career interruption, language, school, working pattern, or social connections. Removing sensitive variables does not guarantee fairness and may make auditing more difficult.

Automation bias and deskilling

Managers may defer to a score to save time or avoid responsibility. Over time, HR professionals may lose investigative and relational skills if they rely excessively on automated recommendations. The appropriate goal is augmented judgement, not the elimination of professional judgement.

Surveillance and trust

Continuous monitoring can change behaviour, intensify work, damage psychological safety, and create a culture of suspicion. A measurable digital trace is not equivalent to productivity, contribution, or wellbeing.

Vendor opacity and concentration

Proprietary systems may limit access to model details, training data, or validation evidence. Vendor concentration can create operational dependency and make migration difficult. Procurement must therefore include transparency, portability, audit, and exit provisions.

Dynamic environments

Work roles, skills, regulations, and labour markets change. A model validated last year may be unsuitable after restructuring, remote-work changes, economic shocks, or new legislation. Continuous review is a substantive requirement, not an administrative formality.

10. The Future of HR Analytics

The future is likely to involve conversational analytics, multimodal work data, real-time skills graphs, AI agents, synthetic data for testing, and increasingly integrated workforce ecosystems. These capabilities may improve accessibility, personalised development, and strategic workforce planning. They may also expand surveillance, blur the boundary between assistance and control, and make decisions harder to contest.

Several principles should guide the future:

• Human-centred design: technology should enhance meaningful work and agency, not merely maximise extraction of labour.
• Skills-based opportunity: capability evidence should be broader than credentials and historical job titles, while remaining verifiable and contestable.
• Privacy by design: protection should be built into architecture, defaults, procurement, and workflows.
• Fairness as continuous practice: fairness requires ongoing measurement, participation, and institutional correction.
• Evidence over novelty: adoption should follow demonstrated value and acceptable risk, not technological fashion.
• Professional stewardship: HR analysts must combine statistical competence with employment law, ethics, organisational behaviour, data governance, and communication.

The future HR analyst will increasingly act as translator, investigator, product steward, risk assessor, and organisational adviser. Technical fluency will remain important, but professional legitimacy will depend on knowing when not to automate, how to explain uncertainty, and how to protect the people represented in the data.

Chapter Summary

AI and digital transformation are reshaping HR through generative AI, people analytics platforms, workflow automation, skills intelligence, and integrated digital employee experiences. Their value depends on clearly defined purposes, valid data, appropriate methods, competent users, and measurable organisational outcomes.

AI-enabled HR is socio-technical. Descriptive, predictive, prescriptive, and generative systems can support decisions but do not eliminate uncertainty or responsibility. Generative AI requires verification because fluent outputs may be inaccurate, biased, or confidential. Skills intelligence can promote internal mobility but must distinguish inferred capability from verified competence.

Responsible HR analytics requires privacy, proportionality, data minimisation, security, fairness, transparency, human oversight, contestability, and accountability. Legal compliance may involve data-protection, employment, equality, accessibility, sectoral, and AI-specific requirements. Standards such as the NIST AI Risk Management Framework and ISO/IEC 42001 can support governance but do not replace legal analysis.

A responsible lifecycle defines the purpose, screens risk, assesses data and alternatives, validates the system, deploys controls, monitors impacts, and reviews or retires the system. The enduring professional challenge is to use technology to improve human decisions while preserving dignity, agency, equity, and trust.

Key Terms

Artificial intelligence: Computational systems that perform tasks associated with human intelligence, such as prediction, classification, language generation, or recommendation.

Automation bias: The tendency to accept an automated output without sufficient independent evaluation.

Consent: A voluntary, informed, specific, and unambiguous basis for agreed data processing.

Data minimisation: Limiting data collection and retention to what is necessary for a specified purpose.

Digital transformation: Organisational redesign enabled by digital technology, data, new processes, and changed operating models.

Disparate impact: A disproportionate adverse effect of a policy or system on a protected or vulnerable group.

Generative AI: AI that produces new content, such as text, code, images, or summaries.

Human-in-the-loop: A design in which a person reviews and approves an AI-supported decision.

Human-in-command: Governance in which people control the purpose, boundaries, deployment, and withdrawal of AI.

Model drift: Decline in model suitability as relationships, populations, or operating conditions change.

People analytics platform: An integrated technology environment for collecting, analysing, and presenting workforce data.

Pseudonymisation: Processing data so direct identifiers are replaced while re-identification remains potentially possible with additional information.

Proxy discrimination: Discrimination arising when apparently neutral variables indirectly represent protected characteristics.

Purpose limitation: Using data only for specified and legitimate purposes.

Selection-rate ratio: The selection rate of a focal group divided by the selection rate of a reference group.

Skills intelligence: Structured and inferred information about workforce capabilities and their relationship to work opportunities.

Socio-technical system: A system whose outcomes emerge from interaction between technology, people, processes, and institutions.

Transparency: Providing understandable information about AI use, purpose, operation, limitations, and outcomes.

Review Questions

Short Answer Questions

1. Distinguish digitisation, digitalisation, and digital transformation.
2. What is the difference between predictive analytics and generative AI?
3. Why is employee consent not always sufficient for workplace monitoring?
4. Define automation bias and explain one way to reduce it.
5. What is meant by human-in-command governance?

Long Answer Questions

1. Evaluate the opportunities and risks of generative AI in recruitment and learning.
2. Explain how privacy, security, fairness, and transparency should be integrated into an HR analytics lifecycle.
3. Discuss why skills intelligence may support inclusion while also reproducing inequality.
4. Compare human-in-the-loop, human-on-the-loop, and human-in-command arrangements.
5. Analyse the relationship between digital transformation, data governance, and organisational change.

Analytical / Application Questions

1. An organisation selects 30% of applicants from a focal group and 50% from a reference group. Calculate the selection-rate ratio and explain why the result is a screening signal rather than definitive proof of unlawful discrimination.
2. Design a governance plan for an AI system that recommends employees for promotion, including risk assessment, validation, human oversight, appeal, and monitoring.
3. A vendor claims that its recruitment model is unbiased because it does not use gender or ethnicity. Critically assess this claim and specify the evidence you would request before deployment.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes digital transformation from digitisation?

A. Digitisation converts analogue information into digital form, whereas transformation redesigns organisational value creation and processes.

B. Digitisation always uses artificial intelligence, whereas transformation uses only spreadsheets.

C. Digitisation concerns strategy, whereas transformation concerns scanning documents.

D. There is no meaningful distinction between the two terms.

Correct Answer: A. Digitisation converts analogue information into digital form, whereas transformation redesigns organisational value creation and processes.

Question 2

What is a principal risk of generative AI in HR?

A. It can produce fluent but inaccurate or unsupported content.

B. It can only process numerical data.

C. It cannot generate text or summaries.

D. It guarantees that historical bias is removed.

Correct Answer: A. It can produce fluent but inaccurate or unsupported content.

Question 3

Which use is generally most suitable for low-risk workflow automation?

A. Automatically dismissing employees based on a risk score

B. Ranking candidates for promotion without review

C. Routing a standard leave request for approval

D. Inferring employee emotions from webcam footage

Correct Answer: C. Routing a standard leave request for approval

Question 4

In a selection-rate ratio, what does the denominator represent?

A. The number of protected characteristics in the dataset

B. The selection rate of the reference group

C. The total number of variables in the model

D. The model's prediction error

Correct Answer: B. The selection rate of the reference group

Question 5

Why can removing gender from a model fail to eliminate discrimination?

A. Models cannot operate without gender data.

B. Other variables may act as proxies for gender or related social conditions.

C. Fairness is achieved only when gender is the sole input.

D. Removing gender automatically creates random decisions.

Correct Answer: B. Other variables may act as proxies for gender or related social conditions.

Question 6

What does human-in-the-loop mean?

A. A person reviews and approves an AI-supported decision before action.

B. A system operates without any human involvement.

C. A vendor alone controls the organisation's AI policy.

D. Employees are excluded from decisions involving AI.

Correct Answer: A. A person reviews and approves an AI-supported decision before action.

Question 7

Which principle requires collecting only data necessary for a specified purpose?

A. Data minimisation

B. Model extraction

C. Automation bias

D. Vendor concentration

Correct Answer: A. Data minimisation

Question 8

Why should inferred skills not automatically be treated as verified competence?

A. Inference may rely on incomplete, unequal, or indirect evidence.

B. Skills can never be represented digitally.

C. Learning records are always illegal to use.

D. Job titles provide perfect evidence of every capability.

Correct Answer: A. Inference may rely on incomplete, unequal, or indirect evidence.

Question 9

Which action is an example of contestability?

A. Preventing employees from seeing any system information

B. Allowing an affected person to correct data and request human review

C. Deleting all audit records after deployment

D. Allowing a vendor to make unreviewable decisions

Correct Answer: B. Allowing an affected person to correct data and request human review

Question 10

What is the main purpose of continuous post-deployment monitoring?

A. To ensure the system never requires updating

B. To detect drift, subgroup harms, security events, and unintended consequences

C. To replace all human oversight with dashboards

D. To increase data collection regardless of purpose

Correct Answer: B. To detect drift, subgroup harms, security events, and unintended consequences