AI risk & governance
When an AI system treats some people or cases unfairly because of one-sided data or design choices.
Formal
A systematic error in which a model's results favour or harm certain groups, usually because the training data under-represents them or reflects past unfair decisions, or because of how the goal was defined.
In plain English
Like a new manager who only ever learned from one old boss's hiring records - they copy every habit, fair or not, and do it at speed.
In practice
A municipality tests an AI tool, trained on years of past case decisions, that ranks unemployed citizens for extra help, and finds it places people from certain areas lower although where they live was never meant to count.
Why it matters
An unfair pattern repeats in every automated decision until someone looks for it, so it can break equal-treatment and data protection law for thousands of people before anyone notices.