Description
Systematic assessment and mitigation of bias in training data that could lead to discriminatory AI system outcomes.
Implementation Guidance
Bias Assessment
Conduct systematic bias assessments on training data: analyze representation across protected attributes (age, gender, ethnicity, disability), measure label distribution fairness, identify proxy variables that could encode bias, and assess historical bias embedded in data sources.
Mitigation Strategies
Apply appropriate mitigation techniques: resampling (oversampling underrepresented groups, undersampling overrepresented groups), re-weighting (adjusting sample weights to balance representation), data augmentation for underrepresented categories, and collection of additional representative data.
Ongoing Monitoring
Bias is not a one-time fix. Monitor production data distributions for drift from training data demographics. Re-assess bias when new data sources are added or when monitoring reveals disparate outcomes.
Evidence Requirements
- Bias assessment reports with demographic analysis
- Mitigation strategy documentation and results
- Pre/post mitigation fairness metric comparisons
- Ongoing bias monitoring dashboards and alerts