## TL;DR
AI systems can amplify societal biases present in training data. Fairness is not a single metric — it requires choosing between competing mathematical definitions that may be mutually exclusive.

## Core Explanation
Fairness definitions: demographic parity (equal positive rates across groups), equalized odds (equal TPR and FPR across groups), individual fairness (similar individuals treated similarly). These cannot all be simultaneously satisfied (Kleinberg et al., 2017 impossibility theorem).

## Detailed Analysis
Bias mitigation strategies span the ML pipeline: pre-processing (reweighting/resampling training data), in-processing (fairness constraints in training objective), post-processing (calibrating decision thresholds per group). Model cards (Google, 2019) and datasheets (Gebru et al., 2018) document model characteristics and limitations.

## Further Reading
- ACM FAccT Conference
- AI Now Institute
- Partnership on AI