Quiz — Chapter 10: Bias in Hiring and HR Systems

Part 2: Bias and Fairness

Instructions: This quiz covers the key concepts from Chapter 10. Multiple choice questions have one correct answer unless indicated. True/False questions should be answered with a brief explanation. Short answer questions should be answered in 3–5 sentences. Applied scenario questions should be answered in 150–250 words.

Part I: Multiple Choice (8 questions, 3 points each)

Question 1. Under the EEOC's Uniform Guidelines on Employee Selection Procedures, the "four-fifths rule" states that adverse impact is indicated when:

A. A protected group's selection rate is less than 40% of the majority group's selection rate B. A protected group's selection rate is less than 80% of the highest-performing group's selection rate C. More than 20% of a protected group's applications are rejected at any single screening stage D. A protected group is selected at a rate lower than its proportion of the applicant pool

Question 2. In January 2021, HireVue discontinued its facial expression analysis feature. What was the company's stated rationale?

A. Independent researchers had published a peer-reviewed study demonstrating the technology caused racial discrimination B. The EEOC had issued a formal determination that facial analysis violated Title VII C. There was a lack of scientific consensus on facial expression analysis, and removing it would improve overall fairness D. The Illinois Biometric Information Privacy Act (BIPA) had been amended to specifically prohibit facial analysis in hiring

Question 3. Amazon's ML-based résumé screening tool, described in Chapter 7 and revisited in Chapter 10, learned to systematically downrank women's résumés because:

A. The engineers explicitly programmed gender as a negative feature B. The training data reflected ten years of hiring in a male-dominated technical environment, and the model learned features correlated with gender C. Women's résumés on average contained fewer of the keywords associated with technical roles D. Amazon's legal team identified women in technical roles as a retention risk and adjusted the model accordingly

Question 4. Which of the following AI hiring tools creates the most direct risk of an Americans with Disabilities Act (ADA) violation?

A. A keyword-based ATS system that filters out résumés lacking certain certifications B. A sourcing algorithm that identifies passive candidates from professional networking sites C. A video interview platform that scores candidates on facial expression patterns without offering alternative assessment pathways for candidates with autism spectrum disorder D. A salary prediction model that calibrates offers based on candidates' historical salary data

Question 5. NYC Local Law 144 (effective July 2023) requires employers using automated employment decision tools to:

A. Obtain explicit written consent from all candidates before using any AI in their evaluation B. Conduct annual independent bias audits and publish results, and notify candidates of AI tool use C. Demonstrate EEOC pre-approval of any AI hiring tool before deployment in New York City D. Replace all AI hiring tools with human-reviewed processes for any role paying over $100,000

Question 6. Under the EU AI Act, AI systems used in employment, worker management, and recruitment are classified as:

A. Minimal-risk systems requiring only transparency disclosures B. Limited-risk systems requiring basic documentation C. High-risk systems requiring risk assessments, human oversight, and conformity assessments D. Prohibited systems that may not be deployed in EU member states

Question 7. The "culture fit" AI assessment problem refers to:

A. AI tools that evaluate candidates' familiarity with the hiring company's products and services B. AI tools that predict cultural fit by comparing candidate profiles to current employees, potentially encoding demographic similarity as a proxy for competence C. The tendency of hiring algorithms to favor candidates from the same geographic region as the company's headquarters D. Screening tools that use cultural background information (language, country of origin) to predict retention

Question 8. Which statement best describes the legal principle governing employer liability for AI discrimination tools?

A. Employers are only liable if they were aware the AI tool was discriminatory before deploying it B. Liability rests with the vendor who designed the tool, not the employer who deployed it C. Employers bear legal responsibility for discrimination caused by AI tools they deploy, regardless of whether the vendor designed the discriminatory feature D. AI tool discrimination claims must be brought against the AI system itself under emerging AI personhood law

Part II: True or False (5 questions, 4 points each — T/F with explanation)

Question 9. True or False: Career gaps on a résumé are a facially neutral screening criterion with no potential for disparate impact on protected groups.

Explain your answer in 2–3 sentences.

Question 10. True or False: Removing candidate names from résumés during initial ATS screening (blind screening) eliminates all forms of demographic proxy bias in automated résumé screening.

Explain your answer in 2–3 sentences.

Question 11. True or False: A vendor's internal validation study showing that its AI hiring tool scores correlate with eventual hire decisions at client companies is sufficient evidence of criterion validity for the tool's deployment in high-stakes hiring.

Explain your answer in 2–3 sentences.

Question 12. True or False: Flight risk prediction tools that do not use gender or disability status as explicit input features cannot produce disparate impact based on these characteristics.

Explain your answer in 2–3 sentences.

Question 13. True or False: Under the Americans with Disabilities Act, an employer can satisfy the reasonable accommodation requirement for an AI video interview by making the accommodation available upon request, without proactively informing candidates that an accommodation option exists.

Explain your answer in 2–3 sentences.

Part III: Short Answer (4 questions, 8 points each)

Question 14. Explain the "prestige bias" problem in ATS keyword screening — how it operates, what protected characteristics it proxies for, and why it constitutes a potential legal issue under Title VII.

Question 15. What is the difference between "disparate treatment" and "disparate impact" under Title VII of the Civil Rights Act, and which is more relevant to AI hiring discrimination? Provide an example of each in the AI hiring context.

Question 16. Describe three specific ways in which HireVue's facial expression analysis could have produced discriminatory outcomes for candidates with disabilities. For each, identify the specific disability, the specific mechanism of disadvantage, and the relevant legal provision.

Question 17. Explain the "gaming problem" in ATS keyword screening — how it emerges, who benefits from gaming knowledge, who is disadvantaged, and what the implications are for the validity of keyword-screened résumés.

Part IV: Applied Scenario Questions (3 questions, 10 points each)

Question 18 — The Audit Finding

You are the Director of Talent Acquisition at a technology company. Your annual adverse impact audit has returned the following results for your ATS screening system:

• White candidates: 52% pass rate through initial screening
• Asian candidates: 49% pass rate through initial screening
• Hispanic candidates: 31% pass rate through initial screening
• Black candidates: 28% pass rate through initial screening
• Women: 44% pass rate through initial screening
• Men: 56% pass rate through initial screening

Analyze these results using the four-fifths rule. Which groups show adverse impact? What are your immediate obligations? What investigations would you undertake to determine the source of the disparities? What would you recommend as next steps to your CHRO?

(Answer in 200–250 words.)

Question 19 — The Vendor Pitch

A video interview AI vendor presents the following at a sales meeting:

"Our platform analyzes 28 behavioral signals from recorded video interviews, including verbal content, speaking pace, and 'engagement indicators.' We have processed over 2 million interviews and have a database of high-performer profiles across 200 client companies. Our system is validated — clients who use our platform report a 23% improvement in 90-day retention and a 15% improvement in first-year performance ratings. We do not use facial expression analysis. Our platform includes an 'alternative assessment pathway' that candidates can request."

Evaluate this pitch from a legal compliance and ethical standpoint. What claims are adequately supported? What claims require further scrutiny? What specific questions would you ask before deciding whether to deploy this platform? What information would you need about the "alternative assessment pathway" before concluding it satisfies ADA requirements?

(Answer in 200–250 words.)

Question 20 — The Executive Decision

You are presenting to your company's executive team on a proposal to adopt AI résumé screening across all 15,000 annual job applications. The CFO argues: "If AI screening is even slightly more efficient than human screening, we should use it — the scale savings justify it, and we can monitor for bias over time." The Chief Diversity Officer responds: "We should not deploy a technology we haven't audited first — by the time bias shows up in diversity metrics, thousands of candidates will have been affected."

Evaluate both positions. What are the strongest arguments for each? What does the "monitor over time" approach require to be ethically defensible? What pre-deployment requirements would bridge the tension between these positions? What recommendation would you make to the executive team, and what conditions would you attach?

(Answer in 200–250 words.)

Answer Key

Part I: Multiple Choice

Part II: True or False

Question 9: FALSE. Career gaps are facially neutral but carry significant disparate impact potential. Women are disproportionately likely to have caregiving-related gaps; veterans have service-transition gaps; individuals with disabilities may have medical gaps. All of these groups have some degree of legal protection. Penalizing gaps without examining their cause systematically disadvantages protected groups.

Question 10: FALSE. Removing names reduces name-based bias but does not eliminate proxy bias. Other demographic indicators remain in résumés: graduation years (proxy for age), school names (proxies for race and socioeconomic status), addresses and zip codes (proxies for race and socioeconomic status), and language patterns. True bias reduction requires removing or deweighting all demographic proxies, not only names.

Question 11: FALSE. Correlating AI scores with hiring decisions demonstrates that the tool predicts who was hired — not that it predicts job performance. If the hiring decisions themselves were biased, then the validation criterion is contaminated. Criterion validity requires demonstrating correlation with actual job performance (ratings, productivity, retention), not with the outcome of a potentially biased prior process.

Question 12: FALSE. Removing explicit features does not prevent proxy discrimination. If variables that remain in the model — communication metadata patterns, badge access data, accommodation request history, tenure patterns — correlate with gender or disability status in the training data, the model will learn to use these correlations to predict departure in ways that indirectly reflect the protected characteristics. This is the proxy variable problem applied to flight risk AI.

Question 13: FALSE (or at minimum legally uncertain and ethically inadequate). EEOC guidance and ADA case law suggest that accommodation must be genuinely accessible, which typically means proactive communication that accommodation options exist — not merely passive availability upon request. Candidates who do not know an accommodation option exists cannot meaningfully exercise the right to it. Best practice and defensible practice both require proactive notification.

Part III: Short Answer — Scoring Guidance

Question 14 (Prestige Bias): A strong answer identifies: (1) how ATS models weight school name and employer prestige, either explicitly or through ML training on historical hires; (2) that elite university enrollment correlates with socioeconomic status, which in turn correlates with race and ethnicity; (3) that this creates a proxy variable relationship — the system is effectively screening on socioeconomic background and race while claiming to screen on academic quality; (4) that under Griggs v. Duke Power Co., a facially neutral practice with disparate impact on a protected group must be justified by demonstrated job-relatedness, which prestige weighting typically is not.

Question 15 (Disparate Treatment vs. Disparate Impact): A strong answer correctly defines both: disparate treatment = intentional discrimination based on protected characteristic; disparate impact = facially neutral practice with disproportionate negative effect on protected group. Notes that disparate impact is more relevant to AI hiring, since AI tools do not typically discriminate intentionally but may produce discriminatory outcomes through biased training data or proxy variables. Provides clear examples: disparate treatment would be an algorithm explicitly told to score women lower; disparate impact would be an algorithm trained on historical data that happened to encode gender bias.

Question 16 (HireVue and Disability): A strong answer identifies three distinct scenarios, for example: (1) ASD — atypical facial affect is coded as low enthusiasm/confidence, mechanism = calibration on neurotypical norms, provision = ADA reasonable accommodation requirement; (2) facial paralysis — candidate cannot produce expressions the system evaluates, mechanism = literal inability to complete the assessment equitably, provision = ADA; (3) anxiety disorder — physiological anxiety symptoms coded as negative signals, mechanism = assessment measuring clinical symptoms rather than job competencies, provision = ADA combined with prohibition on pre-offer medical inquiry.

Question 17 (Gaming Problem): A strong answer explains: (1) ATS optimization advice is available through career coaches, university career services, and online platforms; (2) access to this advice is unequally distributed — candidates with professional networks, career coaches, and well-resourced university career centers have earlier and fuller access; (3) this creates information asymmetry that compounds other screening disadvantages; (4) as optimization becomes widespread, résumés become more homogeneous and less informative, degrading the signal value of language that was formerly distinctive; (5) validity implication: a system optimized-around dataset is not measuring what it was designed to measure.

Part IV: Applied Scenario — Scoring Guidance

Question 18 (The Audit Finding): A strong answer: (1) correctly applies the four-fifths rule — identifies that Hispanic (31/52 = 60%) and Black (28/52 = 54%) candidates show adverse impact, and women (44/56 = 79%) are at the threshold; (2) describes immediate obligations including documentation, escalation to legal counsel, and consideration of whether to pause the tool pending investigation; (3) proposes investigating source of disparity — examining criteria generating the gap, whether keyword criteria have job-relatedness support, whether formatting issues explain any disparity; (4) recommends concrete next steps including tool review, criteria validation, and consideration of additional human review at the threshold.

Question 19 (The Vendor Pitch): A strong answer: (1) notes the retention and performance claims rely on client self-reported data, not independent validation — 90-day retention and first-year manager ratings are not rigorous criterion validity evidence; (2) identifies that "engagement indicators" is undefined and may include facial/behavioral signals warranting clarification; (3) asks what the "28 behavioral signals" are and whether each has independent validity evidence; (4) questions whether the "alternative assessment pathway" is proactively communicated to candidates, whether it is tested for equivalent predictive validity, and whether it has been reviewed by ADA counsel; (5) asks for adverse impact data disaggregated by race, gender, age, and disability status.

Question 20 (The Executive Decision): A strong answer: (1) validates the CFO's efficiency argument while noting that "monitor over time" is inadequate without specified monitoring criteria, response protocols, and defined liability; (2) validates the CDO's concern about harm at scale — if bias operates on 15,000 applications and affects protected groups at even 5%, that is hundreds of candidates per year; (3) articulates that a defensible approach requires pre-deployment adverse impact testing on representative sample data, validity evidence, accommodation planning, and a defined audit schedule with action thresholds; (4) makes a specific recommendation — conditional deployment with pre-deployment testing, structured first-cycle monitoring with a defined threshold for pulling the system, and proactive accommodation pathway — or recommends phased pilot deployment with smaller applicant pool before full rollout.