Algorithmic bias

Algorithmic bias is not only a data problem. NIST SP 1270 groups its sources into three categories: systemic bias embedded in institutions and historical records, statistical and computational bias from unrepresentative samples or flawed measurement, and human bias in how people design, label, interpret and rely on outputs. ISO/IEC TR 24027 describes how bias enters at each lifecycle stage and how to measure it.

Bias matters legally because it can become discrimination. Data protection law restricts processing of special categories of personal data and automated decisions with significant effects, and the EU AI Act requires providers of high-risk systems to examine training data for possible biases and take measures to detect, prevent and mitigate them. Fairness metrics make bias measurable, but choosing which metric applies is a governance decision.

Example: a hiring model trained on past promotions learns to favour candidates from a few universities, reproducing the historical profile of the workforce rather than predicting job performance.