Artificial Intelligence in Sourcing and Recruitment

Introduction

Artificial intelligence (AI) has become an important component of contemporary personnel sourcing and recruitment. Organizations now use algorithmic systems to identify potential candidates, search professional networks, rank applications, generate recruitment content, answer applicant questions, schedule interviews, and estimate likely hiring outcomes. These applications can reduce administrative workload and expand the scale of recruitment. They can also reproduce historical inequalities, obscure decision processes, expose personal data, and create false confidence in predictions.

The central managerial issue is therefore not whether an organization should “use AI,” but under what conditions a particular AI system is valid, explainable, lawful, secure, and subject to meaningful human oversight. Recruitment decisions affect access to employment and income. Consequently, apparently technical choices—such as which data to collect, how to define a successful candidate, or how to measure model performance—are also ethical and organizational choices.

This chapter evaluates major uses of AI in sourcing and recruitment. It distinguishes search, matching, content generation, conversational systems, and predictive tools; examines validity and fairness; and develops a governance approach for responsible implementation.

Learning Objectives

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

1. Explain the main uses of AI across the recruitment process.
2. Distinguish automated search and matching from predictive selection.
3. Evaluate validity, reliability, fairness, explainability, privacy, and security risks.
4. Design appropriate human oversight and governance controls.
5. Interpret recruitment metrics without confusing correlation with causation.
6. Recommend a responsible AI implementation for a sourcing or recruitment problem.

1. AI Across the Recruitment System

AI refers broadly to computational systems that perform tasks associated with human intelligence, including classification, prediction, language processing, optimization, and pattern recognition. Machine learning (ML), a major category of AI, uses data to estimate relationships or patterns rather than relying exclusively on manually written rules. Generative AI produces new text, images, summaries, or other content in response to prompts or structured inputs.

Recruitment AI should be understood as a sociotechnical system. The system includes software, data, vendors, recruiters, hiring managers, applicants, organizational policies, and the labor-market context. A technically accurate model may still produce unacceptable results if its data are biased, its target variable is poorly chosen, or its outputs are treated as authoritative.

A typical recruitment workflow contains five connected stages:

The same candidate may be represented differently at each stage. A professional profile may be suitable for search but insufficient for a selection decision. This distinction is essential: finding a person is not the same as proving that the person is qualified or will perform well.

2. AI-Assisted Search and Sourcing

2.1 Search and recommendation

Traditional sourcing often depends on Boolean searches, recruiter networks, referrals, and job-board databases. AI-assisted systems can expand this process through semantic search. Rather than requiring an exact keyword match, a semantic system attempts to identify related concepts, skills, job titles, qualifications, or career experiences.

For example, a system may connect “process improvement,” “Lean operations,” and “continuous improvement” even when the candidate’s profile does not contain the exact wording in a job description. Recommendation systems may also identify passive candidates whose profiles resemble those of people previously considered for similar roles.

AI-assisted search can improve recall—the proportion of relevant candidates retrieved—but higher recall does not automatically improve hiring quality. A system may return more profiles while also increasing irrelevant results, reproducing network inequalities, or overrepresenting candidates with highly visible online profiles.

Two useful retrieval concepts are:

• Precision: the proportion of retrieved candidates who are relevant.
• Recall: the proportion of all relevant candidates who are retrieved.

If a search returns 40 profiles, of which 20 are genuinely relevant, precision is 20 ÷ 40 = 0.50, or 50%. If the labor market contains 80 relevant candidates and the search finds 20, recall is 20 ÷ 80 = 0.25, or 25%. These measures require a defensible definition of “relevant”; otherwise, numerical evaluation merely disguises subjective judgment.

2.2 Risks in sourcing systems

Sourcing systems can create unequal opportunity in several ways:

1. Visibility bias: candidates with extensive digital footprints may be easier to find than equally qualified candidates with limited online presence.
2. Historical bias: training data may reflect earlier recruitment patterns that favored particular institutions, occupations, regions, or demographic groups.
3. Proxy discrimination: location, school, language, career gaps, or employment history may correlate with protected characteristics.
4. Feedback loops: recruiters repeatedly select recommended candidates, causing future systems to learn that previous recommendations were successful.
5. Taxonomy bias: skill dictionaries may recognize dominant terminology while missing equivalent experience expressed in different cultural or occupational language.

Recruiters should therefore test not only whether a system retrieves plausible candidates, but also who is systematically absent from its results.

3. Matching, Ranking, and Candidate Assessment

3.1 Matching is not selection validity

Matching systems compare job requirements with candidate attributes. They may use rules, embeddings, similarity scores, or supervised learning. A simple conceptual representation is:

Here, S(c,j) is the match score for candidate c and job j; mᵢ(c,j) is the degree of match on attribute i; and wᵢ is the assigned weight. The weights should be non-negative and their interpretation should be documented. This expression is a conceptual weighted-score model, not a universal law. It is appropriate only when attributes can be defined consistently and when the organization has a defensible reason for assigning the weights.

A high match score indicates similarity to a job profile, not necessarily future job performance. Similarity can be useful for prioritizing human review, but the organization must establish job-relatedness through job analysis and validation. A model that predicts whether a candidate resembles previously hired employees may simply reproduce prior hiring preferences.

3.2 Validity and reliability

In personnel selection, validity concerns the extent to which evidence supports interpretations and decisions based on an assessment. Relevant forms include:

• Content validity: whether the assessment represents important job tasks or competencies.
• Criterion-related validity: whether scores relate to a meaningful work outcome, such as standardized performance ratings or sales results.
• Construct validity: whether the tool measures the intended construct rather than an unintended proxy.
• Incremental validity: whether the AI tool adds useful information beyond existing, less intrusive methods.

Reliability concerns consistency. An assessment may produce consistent results and still measure the wrong attribute. Conversely, a potentially valid assessment may be too unstable to support decisions.

A recruitment model should not be judged solely by accuracy. If 95% of applicants do not meet a hiring threshold, a model that labels every applicant “not selected” achieves 95% accuracy while identifying no qualified candidates. Confusion matrices and group-specific error rates are therefore more informative.

Important measures include:

Here, TP, FP, TN, and FN denote true positives, false positives, true negatives, and false negatives. These measures apply to a defined classification task and require a specified threshold. Changing the threshold generally changes the balance between false positives and false negatives. In recruitment, the appropriate balance depends on the purpose, the cost of error, and the consequences for applicants.

3.3 Fairness is multidimensional

No single fairness metric resolves every ethical or legal question. Common concepts include:

• Demographic parity: selection rates are equal across groups.
• Equal opportunity: true-positive rates are equal across groups.
• Equalized odds: true-positive and false-positive rates are equal across groups.
• Calibration: a score has comparable meaning across groups.

These properties may conflict, particularly when groups have different base rates or when predictions are imperfect. Therefore, fairness assessment requires normative judgment, legal advice where appropriate, and consultation with affected stakeholders. Statistical parity alone does not establish that a tool is job-related, and unequal outcomes do not by themselves prove intentional discrimination. Both outcomes and decision processes must be investigated.

4. Generative AI in Recruitment Content

Generative AI can draft job descriptions, outreach messages, interview guides, candidate summaries, rejection communications, and employer-branding material. It can improve speed and help recruiters adapt content for different channels. However, generated content is probabilistic: it may sound authoritative while containing inaccurate, stereotyped, or fabricated information.

Appropriate uses

• Producing an initial draft from an approved role profile.
• Converting technical requirements into accessible language.
• Suggesting inclusive alternatives for unnecessarily restrictive wording.
• Summarizing recruiter notes for human review.
• Generating interview questions tied to validated competencies.

Inappropriate or high-risk uses

• Inventing qualifications, benefits, salary information, or organizational policies.
• Automatically rejecting applicants based on generated summaries.
• Asking a public system to process confidential candidate data.
• Inferring personality, mental health, honesty, or protected characteristics from writing style.
• Sending personalized communications without verifying accuracy and consent.

A responsible workflow labels AI-generated drafts internally, requires human verification, and preserves an audit trail showing the source of factual claims. Prompts and outputs may themselves contain personal or confidential information and must be handled according to data-protection and information-security policies.

5. Chatbots and Conversational Recruitment

Recruitment chatbots can answer routine questions, collect basic information, provide status updates, schedule interviews, and support candidates outside normal office hours. Their value is greatest when they improve access to accurate information rather than simulate human judgment.

A well-designed chatbot should:

1. Identify itself as an automated system.
2. State what information it collects and why.
3. Offer accessible language and alternative communication channels.
4. Escalate complex, sensitive, or disputed matters to a human.
5. Avoid making promises about selection outcomes.
6. Record interactions securely and retain them only as long as justified.
7. Provide a correction route when information is wrong.

Chatbots can disadvantage applicants with disabilities, limited digital access, low literacy, limited proficiency in the chatbot’s language, or unusual career histories. Accessibility testing should include real users and should not be treated as a final cosmetic check.

6. Predictive Tools and Algorithmic Risk

Predictive recruitment tools estimate an outcome from available data. Examples include estimates of candidate response probability, likelihood of completing an application, expected time to fill, or predicted retention. A prediction is not a fact; it is an uncertain estimate conditional on data, assumptions, and a chosen outcome definition.

The most important design question is often the target variable. “Successful employee” might mean performance rating, tenure, promotion, customer satisfaction, or manager preference. Each measure contains assumptions and potential bias. For example, retention may reflect managerial quality, scheduling flexibility, compensation, or discrimination rather than candidate suitability.

Correlation, causation, and managerial misuse

A model may identify that a factor is associated with an outcome without showing that the factor causes it. A candidate’s attendance at a prestigious institution may correlate with prior opportunity and access to professional networks; it does not necessarily cause better job performance. Removing a sensitive attribute from the dataset does not eliminate discrimination if correlated variables remain.

Predictive tools also create risks of automation bias, in which decision makers over-trust a system because it appears objective. A score should therefore be presented with uncertainty, evidence boundaries, and a clear statement of what it must not be used to infer.

7. Explainability and Human Oversight

Explainability has several meanings:

• Technical interpretability: understanding how inputs contribute to an output.
• Procedural transparency: knowing how the tool is used in the workflow.
• Candidate explanation: communicating a comprehensible reason for a decision.
• Contestability: enabling a person to challenge or correct an outcome.

A highly complex model may offer strong predictive performance but weak interpretability. Conversely, a simple model may be easier to explain but less accurate. The appropriate choice depends on the decision’s consequences, the availability of validation evidence, and the feasibility of human review.

Human oversight is not achieved merely by placing a recruiter at the end of an automated workflow. Meaningful oversight requires that the reviewer:

• understands the tool’s purpose and limits;
• can inspect relevant evidence rather than only a score;
• has authority to override the output;
• is not penalized for making a justified override;
• documents significant decisions; and
• can suspend the system when material problems arise.

Human review can itself be biased. Oversight must therefore be structured through job-related criteria, standardized review guides, training, and monitoring.

8. Privacy, Data Protection, and Security

Recruitment systems process identity data, employment histories, education records, assessment results, communications, and sometimes sensitive personal information. Responsible use requires data minimization: collect only what is necessary for a specified purpose. It also requires purpose limitation, accuracy, retention controls, access restrictions, vendor accountability, and secure deletion.

Before deployment, an organization should ask:

• What is the lawful and ethical basis for processing?
• Has the candidate been informed in clear language?
• Is the data necessary for the recruitment purpose?
• Are inferred attributes being created without justification?
• Where are data stored and transferred?
• Can applicants access, correct, or contest relevant information?
• What happens when a vendor uses data to improve its general model?
• How will a breach, erroneous decision, or model failure be reported?

Privacy impact assessments and algorithmic impact assessments should be proportionate to the system’s risk. Contractual controls should address subcontractors, model training, confidentiality, breach notification, deletion, audit rights, and restrictions on secondary use.

Legal requirements differ by jurisdiction. Depending on location and sector, organizations may face employment-discrimination law, data-protection law, accessibility requirements, automated-decision provisions, recordkeeping obligations, and emerging AI-specific regulation. Compliance is a minimum condition, not a substitute for professional ethics or evidence of validity.

9. A Governance Framework for Responsible Recruitment AI

A practical governance model can be organized into six stages.

Stage 1: Define the decision and its consequences

Specify whether AI is supporting search, communication, prioritization, assessment, or a final decision. The more consequential the decision, the stronger the evidence and oversight required. Establish prohibited uses, such as inferring protected characteristics or making unreviewable rejection decisions.

Stage 2: Establish job-related requirements

Conduct job analysis and define essential competencies, performance criteria, and minimum qualifications. Avoid using historical hiring decisions as the only source of “ground truth.” Historical decisions may reflect bias or inconsistent standards.

Stage 3: Evaluate data and model evidence

Document data sources, provenance, representativeness, missingness, labeling practices, and known limitations. Test performance overall and across relevant groups. Examine false-positive and false-negative rates, not merely aggregate accuracy. Use a holdout sample or prospective validation where feasible, and monitor performance after deployment because labor markets and roles change.

Stage 4: Design the human process

Assign responsibilities for reviewing outputs, handling appeals, correcting records, and stopping the system. Provide recruiters with training on automation bias, fairness, privacy, and appropriate interpretation. Ensure that applicants have accessible alternatives and meaningful contact with a human.

Stage 5: Monitor and audit

Track outcomes, complaints, overrides, missing-data patterns, processing times, accessibility incidents, and changes in performance. Monitoring should compare the AI-supported process with a reasonable baseline. An audit should examine both statistical results and organizational practice.

Stage 6: Review, remediate, or retire

A system should be modified or withdrawn when evidence shows unacceptable bias, poor validity, security weaknesses, or unexplained deterioration. Retirement is a governance success when continued use would create greater risk than value.

10. Worked Application: A University Hiring Platform

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

Background

A university wishes to recruit administrative analysts. It receives approximately 2,000 applications for each annual recruitment cycle. The human-resources department proposes an AI tool that ranks applications according to similarity to profiles of employees who received high performance ratings during their first two years.

Problem

The proposal promises faster screening but raises concerns. Historical high performers disproportionately attended a small number of universities, and performance ratings vary across departments. Some applicants have nontraditional careers, caregiving gaps, or international qualifications that are not represented in the training data.

Analysis

The proposed target—historical performance rating—is not automatically a valid measure of future performance. It may contain rater bias and departmental inconsistency. The training sample may also be unrepresentative. A ranking model could reward similarity to the existing workforce rather than job-related competence.

The university conducts a job analysis and identifies competencies including data interpretation, stakeholder communication, attention to detail, and ethical handling of information. It develops a structured application rubric and decides that AI may assist with organizing evidence but may not make an automatic rejection decision.

The university evaluates the tool using historical applications and a prospective pilot. It examines precision and recall, group-specific error rates, accessibility, and the rate at which recruiters override recommendations. It discovers that applicants with international job titles are frequently under-ranked because the system’s taxonomy maps their experience poorly.

Approach and solution

The university adopts the following controls:

• job-related competency definitions approved before model testing;
• exclusion of irrelevant personal variables and high-risk proxies;
• a standardized evidence rubric visible to recruiters;
• human review of all applications meeting minimum qualifications;
• an appeal and correction process for applicants;
• testing across language, disability-access, career-stage, and qualification groups where lawful and ethically appropriate;
• quarterly monitoring and an annual independent review;
• contractual restrictions on vendor data reuse;
• suspension criteria if error disparities or security incidents exceed predefined limits.

The AI system is used primarily to identify evidence and organize applications, not to predict personal worth or “fit.” Recruiters receive training emphasizing that a recommendation is a prompt for review, not a decision.

Outcome

The pilot reduces administrative sorting time, but the organization does not rely on the model’s ranking alone. After improving the international-qualification taxonomy and standardizing performance ratings, the university finds that the tool provides useful assistance for high-volume review. It retains the right to discontinue use if monitoring shows deterioration or if the system cannot provide adequate explanations.

Learning points

The case illustrates five principles. First, validity begins with job analysis. Second, historical outcomes are not neutral labels. Third, human oversight must have authority and evidence. Fourth, fairness testing should examine who is missed, not only who is selected. Fifth, a lower-risk use—organizing evidence—may be preferable to an ambitious predictive use that cannot be justified.

11. Contemporary Developments and Strategic Implications

Large language models are increasing the speed and scale of recruitment content creation, candidate interaction, and résumé summarization. At the same time, synthetic profiles, deepfakes, automated application tools, and AI-generated credentials are complicating identity verification and assessment integrity. Organizations should respond by strengthening structured, job-related assessment rather than expanding intrusive surveillance.

Skills-based hiring is also gaining attention. AI can help translate occupational language into transferable skills and identify adjacent experience. This can support mobility and reduce excessive reliance on prestige signals. Yet skills taxonomies must be governed: they should be reviewed by subject-matter experts, tested for linguistic and cultural coverage, and updated as work changes.

Strategically, HR leaders should treat AI as an augmentation portfolio rather than a single transformation project. Low-risk applications—such as scheduling, search assistance, and draft generation—may deliver value with limited consequences. High-impact applications—such as automated ranking or rejection—require substantially stronger validation, transparency, oversight, and legal review.

Best-Practice Checklist

Before deploying an AI recruitment tool, an organization should be able to answer “yes” to the following questions:

• Is the use connected to a clearly defined, job-related purpose?
• Is there evidence that the tool performs adequately for the intended population?
• Have relevant error rates and subgroup outcomes been examined?
• Are applicants informed and provided with accessible alternatives?
• Can a qualified human review, override, and explain the result?
• Are data collection, retention, security, and vendor practices controlled?
• Is there an appeal, correction, incident-response, and audit process?
• Are monitoring and retirement criteria defined before deployment?

Challenges and Limitations

AI cannot eliminate judgment from recruitment; it relocates judgment into data selection, model design, thresholds, workflow configuration, and interpretation. Fairness metrics can conflict, and demographic information needed for auditing may itself require careful legal and ethical handling. Validation evidence may not transfer across occupations, countries, languages, or organizational cultures. Vendors may provide limited information about proprietary models, making independent evaluation difficult. Finally, human oversight may become symbolic if recruiters lack time, training, authority, or incentives to challenge automated outputs.

The appropriate response is disciplined skepticism rather than technological rejection or uncritical adoption. Organizations should use the least intrusive tool capable of achieving the legitimate objective, demand evidence proportionate to risk, and preserve human responsibility for consequential employment decisions.

Chapter Summary

AI in sourcing and recruitment includes semantic search, candidate recommendation, matching, generative content, chatbots, assessment support, and predictive analytics. Search and matching can expand discovery but may reproduce visibility, historical, taxonomy, and proxy biases. A match score is not proof of job-related validity.

Recruitment AI should be evaluated using job analysis, validity evidence, reliability, precision, recall, subgroup error rates, accessibility, privacy, and security. Fairness is multidimensional; demographic parity, equal opportunity, equalized odds, and calibration answer different questions and may conflict. Generative AI requires fact-checking and confidentiality controls, while chatbots require disclosure, accessibility, escalation, and correction mechanisms.

Meaningful human oversight requires competence, access to evidence, authority to override, documentation, and the ability to suspend a system. Responsible governance defines purpose, validates data and models, designs human processes, monitors outcomes, and retires tools when risks exceed benefits. The strategic goal is not maximum automation but valid, explainable, inclusive, secure, and accountable recruitment.

Key Terms

Artificial intelligence: Computational capability to perform tasks associated with human intelligence.

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

Automation bias: Excessive reliance on an automated recommendation or decision.

Calibration: The extent to which a score has comparable meaning across groups.

Candidate matching: Comparing candidate attributes with job requirements or profiles.

Content validity: Evidence that an assessment represents important aspects of the job.

Demographic parity: Equal selection rates across specified groups.

Explainability: The ability to understand and communicate how an AI output is produced or used.

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

Human oversight: Active, competent, and accountable human review of AI-supported activity.

Machine learning: Methods that estimate patterns or relationships from data.

Precision: The proportion of retrieved or predicted-positive cases that are relevant or truly positive.

Predictive validity: Evidence that an assessment predicts a relevant future criterion.

Proxy discrimination: Use of a variable that indirectly represents a protected characteristic.

Recall: The proportion of all relevant or truly positive cases that are retrieved or identified.

Semantic search: Search based on meaning and related concepts rather than exact keywords alone.

Specificity: The proportion of truly negative cases correctly identified as negative.

Target variable: The outcome a predictive model is designed to estimate.

Validity: The extent to which evidence supports intended interpretations and uses of an assessment.

Review Questions

Short Answer Questions

1. What is the difference between precision and recall in AI-assisted sourcing?
2. Why is a high candidate-match score not necessarily evidence of future job performance?
3. Identify two risks associated with generative AI in recruitment communications.
4. What makes human oversight meaningful rather than merely symbolic?
5. Why should recruitment models be evaluated using group-specific error rates?

Long Answer Questions

1. Explain how historical recruitment data can produce feedback loops and proxy discrimination. Propose controls to reduce these risks.
2. Compare content validity, criterion-related validity, construct validity, and incremental validity in the evaluation of a recruitment AI tool.
3. Discuss the privacy, security, and data-governance requirements for a recruitment chatbot.
4. Explain why fairness metrics may conflict and why no single metric is sufficient for governance.
5. Develop a governance framework for deploying an AI system that ranks applicants for a high-volume vacancy.

Analytical / Application Questions

1. An AI sourcing tool retrieves 60 profiles, of which 24 are relevant. The estimated labor market contains 120 relevant candidates, and the tool retrieves 24 of them. Calculate precision and recall, and interpret the results.
2. A vendor claims that its screening model is “95% accurate.” What additional information should an HR manager request before approving deployment?
3. Using the university hiring case, decide whether AI should be used for automatic rejection, recruiter prioritization, or evidence organization. Justify your choice using validity, fairness, explainability, privacy, and human-oversight criteria.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes recall from precision in AI-assisted sourcing?

A. Recall is the proportion of retrieved candidates who are relevant, while precision is the proportion of all relevant candidates retrieved.

B. Recall is the proportion of all relevant candidates retrieved, while precision is the proportion of retrieved candidates who are relevant.

C. Recall measures model security, while precision measures privacy compliance.

D. Recall and precision are identical measures with different names.

Correct Answer: B. Recall is the proportion of all relevant candidates retrieved, while precision is the proportion of retrieved candidates who are relevant.

Question 2

A recruitment model identifies candidates who resemble employees previously rated as high performers. What is the principal validity concern?

A. Similarity to historical employees may reproduce past preferences rather than measure job-related competence.

B. The model necessarily has perfect predictive validity because it uses employee data.

C. Historical data cannot be used in any recruitment context.

D. The model is valid whenever its overall accuracy exceeds 50%.

Correct Answer: A. Similarity to historical employees may reproduce past preferences rather than measure job-related competence.

Question 3

Which is the most appropriate use of generative AI in recruitment?

A. Automatically rejecting applicants based on generated summaries

B. Inferring personality and honesty from writing style

C. Drafting a job description for human verification against an approved role profile

D. Inventing benefits information to make an advertisement more attractive

Correct Answer: C. Drafting a job description for human verification against an approved role profile

Question 4

Which feature is essential for meaningful human oversight?

A. The recruiter sees only a final score and cannot override it.

B. The system makes decisions without retaining records.

C. The reviewer has training, relevant evidence, and authority to challenge or override the output.

D. The vendor describes the system as objective.

Correct Answer: C. The reviewer has training, relevant evidence, and authority to challenge or override the output.

Question 5

A search returns 40 profiles, 20 of which are relevant. What is its precision?

A. 20%

B. 40%

C. 50%

D. 80%

Correct Answer: C. 50%

Question 6

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

A. Demographic parity

B. Equal opportunity

C. Specificity

D. Calibration only

Correct Answer: B. Equal opportunity

Question 7

Why can a model with 95% accuracy still be useless in recruitment?

A. Accuracy is never a valid statistical measure.

B. It may achieve that result by labeling nearly everyone as not selected when most applicants are unqualified.

C. Accuracy applies only to chatbots.

D. High accuracy proves discriminatory intent.

Correct Answer: B. It may achieve that result by labeling nearly everyone as not selected when most applicants are unqualified.

Question 8

Which practice best reflects data minimization?

A. Collecting every available online attribute about each applicant

B. Collecting only data necessary for a specified recruitment purpose

C. Retaining all applicant data indefinitely

D. Allowing vendors to reuse applicant data without restrictions

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

Question 9

What is a major risk of using retention as the target variable in predictive recruitment?

A. Retention is always unrelated to employment.

B. Retention may reflect managerial quality, working conditions, or discrimination rather than candidate suitability.

C. Retention cannot be measured numerically.

D. Retention automatically reveals protected characteristics.

Correct Answer: B. Retention may reflect managerial quality, working conditions, or discrimination rather than candidate suitability.

Question 10

Which deployment choice is generally lower risk than automatic rejection?

A. Using AI to organize evidence for structured human review

B. Using AI to infer mental health from application language

C. Using AI to reject candidates without explanation

D. Using AI to rank applicants by social-media popularity

Correct Answer: A. Using AI to organize evidence for structured human review