Bias can be mitigated at three points in the ML pipeline: Pre-processing — modify the training data before the model sees it. Remove correlations between sensitive attributes and features, reweight or resample data to balance representation. In-processing — modify the training algorithm itself. Add fairness constraints to the optimization objective, use adversarial training to prevent the model from learning sensitive attributes. Post-processing — modify the model’s predictions after training. Apply group-specific thresholds, calibrate scores differently for each group. Each approach has trade-offs: pre-processing is model-agnostic but may lose information; in-processing is most principled but requires modifying the training loop; post-processing is easiest to implement but treats the model as a black box.

Reweighing: assign different weights to training examples so that the weighted dataset satisfies demographic parity. Overweight underrepresented group-label combinations, underweight overrepresented ones. Resampling: oversample underrepresented groups or undersample overrepresented groups to balance the dataset. Correlation removal (Fairlearn’s CorrelationRemover): apply a linear transformation to remove correlations between sensitive attributes and other features. An alpha parameter controls how aggressively correlations are removed. Data augmentation: generate synthetic examples for underrepresented groups (e.g., using SMOTE or generative models). Label correction: identify and correct biased labels using techniques like confident learning.

Exponentiated Gradient (Fairlearn): the most practical in-processing method. It reduces fair classification to a sequence of standard classification problems. You specify a fairness constraint (demographic parity, equalized odds, etc.) and the algorithm finds the best model that satisfies it. Works with any scikit-learn classifier. Adversarial debiasing (Fairlearn, AIF360): trains two networks simultaneously — a predictor that makes predictions and an adversary that tries to infer the sensitive attribute from the predictions. The predictor is penalized for making predictions that reveal group membership. Constrained optimization: add fairness constraints directly to the loss function (e.g., minimize cross-entropy subject to equalized odds).

Threshold optimization (Fairlearn’s ThresholdOptimizer): the most practical post-processing method. Instead of using the same decision threshold (e.g., 0.5) for all groups, it finds group-specific thresholds that optimize accuracy subject to a fairness constraint. For example, the threshold for Group A might be 0.45 and for Group B might be 0.55, chosen to equalize positive prediction rates. Reject option classification: for predictions near the decision boundary (uncertain cases), override the model’s prediction in favor of the disadvantaged group. Calibration: apply group-specific calibration so that predicted probabilities mean the same thing for all groups (Platt scaling or isotonic regression per group).

Fairness constraints typically reduce overall accuracy — but often less than you’d expect. Research shows: small fairness improvements (reducing disparity from 15% to 10%) often cost less than 1% accuracy. Moderate improvements (reducing disparity from 15% to 5%) typically cost 1–3% accuracy. Perfect parity can cost 5–10% accuracy, depending on how different the base rates are. The Pareto frontier maps all achievable fairness-accuracy combinations. Fairlearn’s dashboard visualizes this frontier, helping you choose the right trade-off point. The key insight: you’re not choosing between “fair” and “accurate” — you’re choosing a point on a spectrum.

LLMs require different debiasing approaches: RLHF (Reinforcement Learning from Human Feedback) — train the model to prefer outputs that human raters judge as unbiased. This is how ChatGPT, Claude, and Gemini are aligned. Constitutional AI (Anthropic) — define a set of principles (“constitution”) and train the model to follow them, reducing reliance on human raters. Prompt engineering — include debiasing instructions in the system prompt (“Evaluate candidates based only on qualifications, not names or demographics”). Output filtering — apply guardrails to detect and block biased outputs. Fine-tuning on balanced data — fine-tune the model on a carefully curated, balanced dataset. Representation engineering — identify and modify the internal representations that encode bias.

Fairness gerrymandering: a model can satisfy fairness constraints for predefined groups while still being unfair to subgroups (e.g., fair for women and fair for Black people, but unfair for Black women). Metric gaming: optimizing for one fairness metric can worsen another (impossibility theorem). Fairness washing: using fairness tools superficially to claim compliance without genuine commitment to equity. Ignoring intersectionality: most fairness tools only handle one sensitive attribute at a time; real people belong to multiple groups simultaneously. Static fairness: achieving fairness at deployment doesn’t guarantee fairness over time (feedback loops, distribution shifts). Over-reliance on technical fixes: bias mitigation algorithms can’t fix fundamentally flawed data or unjust systems.

Step 1: Define — identify protected groups and choose the appropriate fairness metric for your context (involve stakeholders). Step 2: Measure — compute disaggregated metrics and fairness scores using Fairlearn or AIF360. Establish a baseline. Step 3: Mitigate — start with post-processing (ThresholdOptimizer), then try in-processing (ExponentiatedGradient) for better trade-offs. Step 4: Evaluate — check that mitigation improved the target metric without unacceptable accuracy loss. Check for fairness gerrymandering (subgroup analysis). Step 5: Document — create a model card documenting the fairness analysis, chosen metric, trade-offs, and limitations. Step 6: Monitor — track fairness metrics in production over time. Feedback loops can erode fairness after deployment.