ai-enabled-hr-talent-automation

Fairness and Bias Governance

Title: Fairness and Bias Governance

Version: 1.0

Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead]

Status: Draft

Last reviewed: 2026-09-07

Next review: [TENANT_CONFIGURATION_REQUIRED]

Reviewers: AI Governance, HR, Legal, Security

Purpose and scope

Defines fairness requirements for AI-assisted candidate matching and the process for testing and reviewing for bias. Definitions of “protected characteristic” vary by jurisdiction — [LEGAL_REVIEW_REQUIRED] for the authoritative list per deployment region.

Excluded characteristics (default — extend per jurisdiction, [LEGAL_REVIEW_REQUIRED])

Age, gender, religion, caste, disability, marital status, ethnicity, photo, nationality, family status, and any other characteristic unrelated to job-relevant qualifications.

Matching design rules

• CV extraction schema does not include fields for the excluded characteristics above; if source documents contain photos or such data, extraction skills must not surface them to the matching skill.
• Matching criteria are limited to mandatory/preferred JD criteria explicitly authored by the requester (skills, experience, certifications) — see candidate-matching-system-prompt.md.
• Match rationale must reference specific job-relevant criteria, not holistic/unexplained scores — supports explainability and auditability of fairness.
• Matching weights are configurable but validated against a schema that rejects any weight tied to an excluded characteristic (config/schemas/rag-config.schema.json / matching-weight config — enforcement point: [TENANT_CONFIGURATION_REQUIRED] validation rule).

Fairness testing process

flowchart LR
    A[Curate test candidate set with varied, de-identified profiles] --> B[Run matching against fixed JD]
    B --> C[Compare score distributions across profile variants differing only in excluded characteristics]
    C --> D{Statistically significant disparity?}
    D -- Yes --> E[Block release, investigate root cause]
    D -- No --> F[Approve for release, log evaluation result]

See ai-evaluation-strategy.md for how this integrates into the release pipeline, and ai-evaluation-scorecard.md for scorecard fields.

Review cadence

Fairness testing runs on every prompt/model version change affecting the matching skill, and on a configurable periodic cadence (default: quarterly — [TENANT_CONFIGURATION_REQUIRED]) even without a change, to catch drift.

Human oversight

Regardless of fairness testing results, every shortlist decision requires human approval (see human-approval-matrix.md); AI recommendations are advisory inputs, not determinative outputs.

Escalation

Any detected bias incident is handled per incident-response-runbook.md and logged in model-risk-register.md.

Change control

Version	Date	Author	Change
1.0	2026-09-07	Documentation package generation	Initial creation

This site is open source. Improve this page.