Concepts & cohorts

• Sensitive attributes, proxies, and context-specific cohorts (region, language, device, channel).
• Document assumptions and intended beneficiaries; declare excluded cohorts with rationale.

Metrics & parity

• Selection rate parity, equal opportunity, calibration, error rate parity, disparate impact (domain-appropriate).
• Define harm thresholds and intervention triggers; link to governance waivers.

Datasets & labels

• Representativeness audits; labeler guidance; leakage checks; licensing & lawful basis documentation.

Mitigation strategies

• Pre-processing (rebalancing, de-biasing); in-model constraints; post-processing calibration and overrides.

Governance & waivers

• Who can approve trade-offs; how long waivers last; re-evaluation cadence; documentation requirements.

PMM signals

• Live parity dashboards; alerts on drift; cohort expansion triggers; automated incident opening on breach.

Documentation & evidence

• Bias evaluation reports; data cards; decision logs; user studies; A/B archives; screenshots of parity trends.

Bias red-teaming

• Prompts designed to elicit unfair behaviour; scripted scenarios; counter-speech and refusal patterns.

User impacts & redress

• Appeals and corrections; explanations; human review; tracking time-to-resolution and outcomes.

Regulatory expectations

• EU AI Act risk management; equality laws; UK ICO fairness guidance cross-references.

Common pitfalls

• One global metric; missing minoritised cohorts; unreviewed trade-offs; untested real-world drift.

Fairness checklist

• Cohorts defined; metrics & thresholds live; mitigations tested; PMM alerts wired; user redress in place.

© Zen AI Governance UK Ltd • Regulatory Knowledge • v1 05 Nov 2025 • This page is general guidance, not legal advice.