Also known as: algorithmic bias
In plain English
AI learns from data created by people, so it can pick up their biases too. A hiring tool trained on past hires might unfairly favour the kinds of people who were hired before.
In practice
Bias is a legal, ethical and reputational risk, especially in hiring, lending, insurance and public services. Test outcomes across different groups before launch, keep monitoring them after, and document how you checked.
Under the hood
Bias can enter through unrepresentative data, historical inequities in labels, proxy variables, model design or the deployment context. Fairness metrics include demographic parity, equalised odds and calibration across groups; they can conflict, so choosing one is a policy decision as much as a technical one.
Example
"The audit found bias in how the model scored applicants from regional areas."