Systematic skew in data, labels, or usage that produces unfair outcomes—fairness goals must be explicit and measured, not assumed.
Bias can enter via historical data, sampling, proxy variables, or feedback loops. Fairness is not one number: different metrics trade off differently across groups and time.
Product and engineering
Define harms, measure, monitor drift after deploy, and document trade-offs. Pairs with responsible AI and regulation awareness such as the EU AI Act for high-risk contexts.