Obizworks-HR
Quarterly Bias Audit

Bias Audit Results

Every three months we run a four-fifths-rule analysis on our AI screening to check that selection rates across inferred demographic groups stay within fair bounds. The full results are published here.

Current period: Q1 2026 (Jan–Mar) · Published 15-Apr-2026 · Next audit due 15-Jul-2026

Q1 2026 Headline
PASS all monitored groups

All inferred demographic groups achieved selection-rate ratios above the four-fifths (0.80) threshold against the highest-selected group. No flags triggered for manual review this quarter.

Q1 2026 at a glance
Candidates screened 2,184 across 47 roles
Shortlisted by AI 348 15.9% overall rate
Groups monitored 6 2 gender · 4 age bands
Lowest ratio 0.87 above 0.80 threshold

Four-fifths analysis

For each monitored group we compute selection rate = (candidates shortlisted by AI) ÷ (candidates screened). The four-fifths rule requires that the lowest group's selection rate is at least 80% of the highest group's rate. Methodology ↓

Inferred gender

Pass · ratio 0.91
Group Screened Shortlisted Selection rate Ratio vs highest
Inferred male 1,412 232 16.4% 1.00 (highest)
Inferred female 772 116 15.0% 0.91

Inferred age band (from graduation year)

Pass · lowest ratio 0.87
Group Screened Shortlisted Selection rate Ratio vs highest
Under 30 682 102 15.0% 0.92
30–39 948 154 16.2% 1.00 (highest)
40–49 412 62 15.0% 0.93
50+ 142 20 14.1% 0.87

Intersectional (gender × age) where n≥100

Pass · lowest ratio 0.84
Group Screened Shortlisted Selection rate Ratio vs highest
Male · 30–39 624 104 16.7% 1.00 (highest)
Male · Under 30 438 68 15.5% 0.93
Female · 30–39 324 50 15.4% 0.92
Female · Under 30 244 34 13.9% 0.84

Intersectional cells with fewer than 100 candidates are excluded to avoid noise; we note their existence and review them descriptively but don't apply the four-fifths test until the sample is large enough.

Historical trend

Lowest four-fifths ratio observed in each quarter. We publish whether we passed and what the lowest ratio was.

Q3 2025 flag — what happened and what we did

In Q3 2025, the female-Under-30 cell came in at 0.79 selection-rate ratio — just below the four-fifths threshold. Root-cause review traced this to two job descriptions that used heavily masculine-coded language ("dominant force", "ninja", "rockstar"). After we asked clients to revise those JDs (with neutral re-writes from our recruiters), the next quarter cleared the threshold at 0.85.

We added a JD language check to our intake flow in Oct 2025 to catch this before screening starts. Full root-cause memo (Q3 2025) ↓

Methodology

What we measure

For each candidate screened in the quarter, we record (a) the AI score and (b) whether they were shortlisted by AI (score ≥ threshold for that role). We then group candidates by inferred demographic attributes and compute selection rates and ratios.

How we infer groups (never used as AI inputs)

  • Gender: inferred from first name using a multi-region name-frequency lookup. Names with <70% gender confidence are excluded.
  • Age band: inferred from earliest graduation year on the CV. Candidates with no graduation year are excluded.

These inferences are used only for audit. They are not provided to the screening AI. We retain inferred groups in anonymised form (no PII) for trend analysis.

Threshold & remediation

A four-fifths ratio below 0.80 triggers (i) manual review by the governance program, (ii) root-cause investigation, (iii) corrective action if a cause is identified, and (iv) public publication of the root cause and remediation alongside the audit results.

Adversarial pre-deployment testing

Before any model change goes live, we run a synthetic candidate set with varied implicit signals (names, locations) and verify that the AI’s scoring isn’t correlated with those signals. See AI Disclosure §7.

Limitations

  • Inferred gender is binary by current methodology — this misrepresents non-binary candidates. We are exploring better signals.
  • Age band is inferred from graduation year, which may be wrong for career-changers or self-taught candidates.
  • Sample sizes for intersectional cells can be small; we publish the cells but acknowledge confidence intervals are wide.

Concerns about bias?

If you believe AI screening at Obizworks-HR has produced an unfair outcome — for you or for a group — raise it through our right-of-explanation channel, or directly to [email protected]. Every concern is reviewed and feeds into the next quarterly audit.