Algorithmic bias

Algorithmic bias occurs when an AI system produces results that systematically disadvantage certain groups (by sex, age, origin, disability…). It is rarely intentional: AI learns patterns from data, and if the data reflects past inequalities or does not represent everyone well, the model repeats or amplifies them.

A classic example: a recruitment system trained on the CVs of people hired over the last ten years may learn to penalise profiles that were rarely hired before, even though nobody asked it to.

The AI Act requires examining possible biases in the training data of high-risk systems and taking measures to detect and correct them (Article 10). Common techniques are measuring fairness metrics per group, reviewing data representativeness and adversarial testing (red teaming).