Bias and Fairness in AI Systems

Bias enters AI systems at multiple points: in data collection (who is represented), in labeling (who labels and how), in problem framing (what is being optimized), in model architecture, and in deployment context.

Most documented harms — facial recognition errors on darker-skinned faces, biased risk scores in criminal sentencing, gender bias in hiring screens — trace to combinations of these factors rather than a single cause.

There are mathematical fairness criteria — demographic parity, equalized odds, calibration — and they are provably incompatible in general. Choosing among them is a value judgment, not a technical one.

Practical fairness work involves disaggregated evaluation, model cards, and red-teaming across protected attributes. Tools like Fairlearn and Aequitas provide standardized measurements.

Pre-processing approaches rebalance training data; in-processing approaches modify model objectives; post-processing approaches adjust predictions. No technique fully eliminates bias; each shifts where it appears.

The most effective mitigation is upstream: better data, clearer problem definitions, and inclusive design teams. Technical fixes alone routinely fail.

The EU AI Act categorizes hiring, credit, and law enforcement AI as high-risk, requiring bias testing. US sector-specific rules (EEOC, CFPB) apply existing civil rights frameworks to AI-mediated decisions.