AI Bias and Fairness Monitoring
Continuously monitor AI system outputs for discriminatory patterns across protected demographic attributes in production.
Objective
Detect and remediate disparate impact in AI decision-making before it causes regulatory exposure or harm to affected individuals.
Maturity Levels
Initial
No bias monitoring exists; disparate impact is not measured.
Developing
Bias was assessed pre-deployment but no ongoing monitoring is in place.
Defined
Automated monitoring tracks outcome distributions across protected attributes with defined alert thresholds.
Managed
Monitoring results are reviewed by a designated fairness owner; disparate impact findings trigger investigation.
Optimizing
Fairness metrics are included in model deployment gates; monitoring methodology is externally validated.
Evidence Requirements
What an auditor or assessor would expect to see for this control.
- —Fairness metrics documentation defining which demographic groups are monitored and which metrics are used (e.g., demographic parity, equalized odds)
- —Periodic bias monitoring reports showing metric values by group over a defined period
- —Disparity alert records for any group difference exceeding defined thresholds, with investigation and resolution
- —Remediation records for confirmed bias findings, including actions taken and post-remediation metric results
- —Legal or compliance review records confirming the monitoring approach satisfies applicable anti-discrimination obligations
Implementation Notes
Key steps
- Define which fairness metrics apply to your use case before deploying — demographic parity, equalized odds, and individual fairness make different trade-offs and are not simultaneously achievable.
- Identify which protected attributes are relevant to each AI use case and ensure you have sufficient data to detect disparate impact across those groups.
- Establish a threshold for 'material disparity' in advance — without a pre-defined threshold, remediation decisions become subjective and difficult to defend.
- Report bias monitoring results to governance bodies regularly; bias that is known and unaddressed creates significant regulatory and legal risk.
Example Implementation
Consumer lending company monitoring an AI model that influences credit limit decisions
Bias and Fairness Monitoring Dashboard — Credit Limit Model
Protected attributes monitored: Gender, race/ethnicity (proxied via zip code), age group (18–25, 26–45, 46–65, 65+)
Fairness metric: Demographic parity — difference in approval rate between groups
Monthly monitoring report (sample):
| Group | Approval Rate | Vs. Reference Group | Material Disparity? |
|---|---|---|---|
| Male | 63.2% | Reference | — |
| Female | 61.8% | -1.4 pp | No (threshold: > 5 pp) |
| Age 18–25 | 54.1% | -9.1 pp | Yes — investigate |
| Age 26–45 | 66.3% | +3.1 pp | No |
| Age 46–65 | 64.2% | +1.0 pp | No |
Materiality threshold: 5 percentage point difference from reference group
Action taken for Age 18–25 flag: Opened investigation to determine if disparity is explained by legitimate credit risk factors or represents proxy discrimination — Legal and Compliance notified per policy
Governance reporting: Results presented to AI Risk Committee monthly
Control Details
- Control ID
- MON-003
- Domain
- Monitoring & Drift
- Typical owner
- AI Governance Team / Legal / Compliance
- Implementation effort
- High effort
- Agent-relevant
- No