In plain English
Bias is systematic error in an AI system's outputs that produces unfair or discriminatory results. It usually enters through training data that reflects historical inequalities, through design choices that suit some groups better than others, or through evaluation that overlooks parts of the population. Because models learn from the past, they tend to reproduce and even amplify the patterns in it unless that is actively checked for.
Why it matters
Biased AI in hiring, lending, policing, or healthcare can scale unfair outcomes far faster than any individual decision-maker, which is why detecting and mitigating it is central to responsible AI.
A worked example
A CV-screening tool trained mainly on a company's past hires can learn to favour the profiles that were historically hired, quietly disadvantaging strong candidates who do not fit that mould.
Common confusion
Bias is not always deliberate, and removing obvious attributes like names does not remove it. Models can pick up bias indirectly through correlated signals such as postcode or schooling.