HR pilots an AI screen for city job applications. Which audit is most appropriate before wider use?

Before wider use of an AI screen for city job applications, the most appropriate audit is a fairness review for disparate impact across applicant groups. Civic hiring tools can encode or amplify bias from historical data, proxy features, or uneven error rates, creating civil-service and legal risk. Measuring selection rates, score distributions, and error patterns across protected or policy-relevant groups—and documenting mitigation—is the assurance step that belongs in a pilot gate. HR and equity stakeholders need evidence the screen does not systematically disadvantage qualified applicants before scale-up across departments.

Auditing the pilot for predictive parity across applicant groups fails the gate: equal precision among those flagged as qualified does not show whether selection rates themselves disadvantage a protected or policy-relevant group. Auditing the pilot for individual fairness across applicant groups fails the gate: requiring similar individuals to receive similar scores depends on a contested similarity metric and does not provide the group-level selection-rate evidence needed for a hiring pilot. Auditing the pilot for overall accuracy across applicant groups fails the fairness question: high overall accuracy can still mask disparate impact across protected or policy-relevant groups.

Exam caveat: Disparate-impact analysis is necessary but not always sufficient; also examine feature legitimacy, human override paths, and adverse-action processes. An individual-fairness review answers a different fairness question than disparate-impact analysis and cannot substitute for it at the pilot gate, because it does not measure whether selection rates disadvantage a protected or policy-relevant group. Distinguish bias audits from accuracy audits—both matter for hiring AI.

Operational check: Run a documented disparate-impact analysis on the pilot cohort before expansion; involve HR/legal; keep applicant scores confidential; define mitigation and human-review paths for borderline cases; and re-audit after model or data updates—not merely after predictive-parity analysis across applicant groups or an individual-fairness review across applicant groups looks healthy.