Bias (in AI)

Systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others.

How it works

Bias in AI systems originates primarily from training data that over- or under-represents certain groups, labels that encode historical human prejudice, and objective functions that optimise for aggregate performance while ignoring disparate impact on subgroups. The model learns statistical associations in the training data — including the prejudiced ones — and applies them to new inputs. Bias can also be amplified by feedback loops where a biased model generates biased outputs that become future training data.

Why it matters

Biased AI systems can cause real harm at scale: loan approval models that discriminate by race, hiring tools that penalise women, or facial recognition systems with high error rates for dark-skinned individuals. Because AI decisions are often automated and opaque, affected individuals may have no recourse. Detecting and mitigating bias is therefore both an ethical imperative and an increasingly legal requirement as AI regulation matures worldwide.