bias mitigation

Bias mitigation reduces unfair biases in model training, data, and outputs affecting demographic groups. **Bias types**: Representation (training data imbalance), association (stereotypical correlations), selection (biased data collection), measurement (inconsistent labeling). **Training-time mitigation**: Data augmentation to balance representation, counterfactual data augmentation, adversarial debiasing (train to be invariant to protected attributes), fair loss functions. **Inference-time mitigation**: Output re-calibration across groups, filtered decoding to avoid stereotypes, prompt-based steering. **Data approaches**: Audit training data for representation, remove biased correlations, collect from diverse sources. **Evaluation**: Test across demographic slices, use fairness benchmarks (BBQ, WinoBias), red-teaming for bias. **Challenges**: Defining "fair", intersectionality, lack of demographic labels, cultural variation in bias. **Transparency**: Document known biases, model cards, intended use guidelines. **Trade-offs**: Fairness metrics can conflict, may reduce overall accuracy, requires ongoing monitoring. **Best practices**: Continuous evaluation, diverse evaluation teams, stakeholder input. Essential for responsible AI deployment.