Overview

Lecture URL: https://skill.re/learn/recruiting/auditing-ai-assisted-decisions-sampling-methodology-fairness-metrics.php

TRANSCRIPT: Auditing AI-Assisted Decisions: Sampling Methodology and Fairness Metrics

Course: AI for Recruiters - Professional Credential

Module: Level 4: Workflow Integration

Section: Chapter 18 -- Quality Systems for AI-Assisted Hiring

Theme: Quality Systems for AI-Assisted Hiring

Lecture: 18.4

Duration: 90 min

Format: Workshop + Case Studies

Audience: Senior recruiters, team leads, recruiting managers

Prerequisites: L3 Certification

What you will learn: Design and conduct audits of AI-assisted hiring decisions. Learn sampling methodology to audit efficiently at scale. Understand how to calculate fairness metrics and detect bias. Create audit systems that run continuously, not just annually.

When AI makes hiring decisions, you need confidence that those decisions are fair and consistent. The way you gain that confidence is through auditing--systematic review of AI decisions to verify they're working as intended.

Auditing AI is different from auditing human decisions. With human decisions, you check that criteria were applied. With AI decisions, you need to verify both that the AI applied its own criteria consistently and that the AI's criteria don't have hidden biases.

In this session, you'll learn to design audit systems that are practical at scale. You can't manually review every decision, so you'll learn sampling methodology that gives you confidence without reviewing everything. You'll learn how to calculate fairness metrics that reveal whether AI has disparate impact. And you'll learn to operationalize auditing so it happens continuously, not just annually.

AUDIT SCOPE: WHAT YOU'RE CHECKING

An effective AI audit checks several things:

Decision consistency: Does the AI make the same assessment of similar candidates? If you run the same resume through the AI twice, does it score the same? If not, there's a problem (though minor variance might be acceptable if the AI uses randomization).

Criteria adherence: Is the AI using the criteria it's supposed to use? If the AI is supposed to score candidates on "relevant experience" and "technical skills," are those the actual drivers of the score? Or is it using proxies like "prestigious company" or "uncommon name"?

Fairness: Does the AI have disparate impact--screening out certain demographic groups at higher rates than others? Does it treat candidates from different backgrounds differently?

Output accuracy: When the AI rates someone as high-potential, are they actually high-potential? Or is the AI misleading you?

SAMPLING METHODOLOGY: AUDITING AT SCALE

You can't audit every decision, but you can audit a representative sample.

Stratified random sampling: If you screen 1,000 candidates per month across five roles, you might audit 100 (ten percent). But to ensure fairness, stratify by demographics: ten percent of women, ten percent of men, ten percent of candidates from different locations, etc. This ensures your sample is representative and you catch any demographic-based disparities.

Risk-based sampling: Audit more aggressively for high-risk decisions. For example, audit every AI rejection where the score was close to the advancement threshold (maybe the AI scored them 48 on a 0-100 scale and 50 is the threshold). These borderline cases are where bias is most likely to matter.

Continuous sampling: Don't audit once per year. Audit weekly or monthly on a rolling basis. This lets you catch problems early rather than discovering them in an annual audit.

Example:

Total AI decisions this month: 500

Sample size: 50 (10 percent)

Stratify by: gender, location, education background

Risk stratification: include ALL decisions where score was within 10 points of threshold

Result: you audit a mix of routine decisions, borderline decisions, and a representative demographic sample

FAIRNESS METRICS: QUANTIFYING BIAS

Fairness audits measure whether AI has disparate impact. The fundamental question: Do candidates from different groups advance at different rates?

Key fairness metrics:

Disparate Impact Ratio (DIR): For each demographic group, calculate the advancement rate. Example: 40 percent of men advance, 25 percent of women advance. The DIR is 25/40 = 0.625. A DIR of less than 0.8 (four-fifths rule) is considered evidence of disparate impact.

Separate rates by stage: Don't just measure overall advancement. Measure advancement at each stage: screen->phone screen, phone screen->interview, interview->offer. Bias might appear at one stage and not others.

Predictive validity by group: For each demographic group, measure: does the AI's scoring actually predict job performance? If the AI scores men accurately but women inaccurately, that's a problem.

Coverage: What percent of candidates from each demographic group are screened by the AI? If the AI only processes applications from certain locations, it might have different coverage for different demographics.

CONDUCTING A FAIRNESS AUDIT: STEP-BY-STEP

Here's a practical approach to conducting an audit.

Step 1: Define your demographic categories. Usually these include gender, race/ethnicity, age, and education background. Define them clearly and consistently.

Step 2: Collect data. For your sample of AI decisions (or all decisions if sample is small), collect: demographics, AI score, AI decision (advance/reject), human reviewer assessment, final outcome.

Step 3: Calculate advancement rates by group. For each demographic group, what percent advanced at each stage?

Step 4: Calculate DIR. For each stage, calculate whether any group has DIR < 0.8.

Step 5: Investigate disparities. If you find disparate impact, investigate. Is the AI's scoring legitimate but just captures real differences in qualifications? (Example: if certain schools produce better candidates and correlate with demographic group.) Or is the AI using biased proxies? (Example: the AI is rating candidates from certain locations lower.)

Step 6: Document findings. Write down what you found, what might explain it, and whether you need to adjust the AI or collect more information.

Step 7: Determine action. If disparate impact is found, what will you do? Retrain the AI? Add constraints to prevent bias? Use human override for certain decisions?

Anti-Pattern 1: Auditing Without Action

An organization conducts an audit and finds that AI screening has disparate impact--women advance at 20 percent but men at 35 percent. The finding is documented. But nothing changes. The AI continues making biased decisions.

Why it happens: Finding bias is uncomfortable. Acting on it requires investment and change.

What goes wrong: You discover a problem but don't fix it. Fairness doesn't improve. Legal risk increases.

How to avoid it: Before auditing, commit to acting on findings. If the audit reveals a problem, you'll fix it.

Anti-Pattern 2: Audit Sampling That Misses Bias

An organization audits AI decisions by reviewing a random sample. They find no disparities. But because their sample was small and skewed (mostly cases where the AI made obvious decisions), they missed borderline cases where bias appears. Bias exists but the audit didn't catch it.

Why it happens: Random sampling can miss patterns, especially if the pattern occurs in a subset of decisions.

What goes wrong: You audit, find nothing wrong, and trust the AI when it's actually biased. You miss an opportunity to improve.

How to avoid it: Use risk-based sampling. Oversample borderline decisions where bias is most likely to matter. Stratify by demographics to ensure you're checking for group-level disparities.

Anti-Pattern 3: Measuring Intent Rather Than Impact

An AI system is designed to screen based on skills. In auditing, the organization checks: "Does the AI scoring correlate with our defined skills?" It does. So they conclude the AI is fair. But they don't check: "Does the AI have disparate impact on demographic groups?" When they finally check, they find disparate impact. The AI was using skills-based criteria (good intent) but achieving biased outcomes (bad impact).

Why it happens: It's natural to focus on whether the AI is using the criteria it's supposed to use, rather than whether the actual outcomes are fair.

What goes wrong: You optimize for the wrong dimension. You build an AI that's consistent but biased.

How to avoid it: Always measure impact, not just intent. An AI that intends to be fair but achieves disparate impact is still a problem.

[PRACTICE PROMPTS]

1. If you were to audit AI decisions in your recruiting, what would you want to check? Design an audit plan that answers your key questions about fairness and accuracy.
2. Calculate a simple fairness audit for a role you're currently hiring for. For your last 50 applications, calculate advancement rates by demographic group at each stage. Do you see disparities?
3. Design a sampling strategy for auditing AI decisions at scale. If you make 1,000 decisions per month, what size sample would you audit? How would you stratify to ensure fairness check?
4. Create an audit report template. What questions would it answer? What data would it include?
5. If you discovered that an AI tool had disparate impact, what would you do? Create a decision-making framework for responding to bias findings.
6. Audit AI decisions to verify they're fair and consistent. Don't assume AI is objective; check.
7. Use stratified random sampling to audit efficiently at scale. Oversample borderline decisions where bias is likely to appear.
8. Calculate fairness metrics (disparate impact ratios) comparing advancement rates across demographic groups. DIR < 0.8 indicates disparate impact.
9. Measure impact, not just intent. An AI that intends to be fair but achieves disparate impact is still a problem.
10. Audit continuously, not just annually. Monthly or quarterly audits let you catch problems early.
11. When you find problems, document them and commit to fixing them. Auditing without action wastes effort.

[GLOSSARY]

Disparate Impact: A hiring practice that appears neutral but disproportionately affects members of a protected group. Measured as Disparate Impact Ratio (DIR).

Disparate Impact Ratio (DIR): The advancement rate of the protected group divided by the advancement rate of the non-protected group. DIR < 0.8 indicates potential disparate impact.

Stratified Sampling: Dividing the population into groups and sampling from each group proportionally to ensure the sample is representative.

Risk-Based Sampling: Prioritizing audit of higher-risk decisions (e.g., borderline cases) where bias is more likely to matter.

[SYNTHESIS AND APPLICATION]

Auditing is how you maintain confidence that AI is working fairly. Build auditing into your normal cycle, not as a special event. Regular, ongoing auditing is better than annual audits.

[REFLECTION EXERCISE]

1. What concerns you most about AI fairness? What metric would give you confidence that the concern isn't happening?
2. If you discovered that your AI had disparate impact, how would you feel? How would you explain it to stakeholders?
3. What's the fairness question most important to your organization? Design an audit to answer that question.
4. How often should you audit? Weekly, monthly, quarterly, annually? Why?
5. If two audits gave conflicting results about fairness, how would you investigate which was more trustworthy?

[CLOSING REMARKS]

Auditing AI is an investment in fairness and defensibility. It's not optional; it's essential.

AI for Recruiters Certification Program

Level 4: Workflow Integration | Quality Systems for AI-Assisted Hiring | Lecture 4

A SkillsClinic initiative.

Duration: ~90 minutes | Word Count: ~2,350