GLOSSARY

A systematic pattern in an algorithmic system that can produce unfair, inaccurate, or disproportionately harmful outcomes for particular people or groups.

Unmanaged bias can create legal exposure, exclusion, poor decisions, reputational damage, and loss of trust. Identifying it requires stakeholder analysis, representative evaluation, documented trade-offs, and monitoring after deployment.

A hiring model is evaluated by role and relevant demographic groups. The organisation discovers that a proxy feature disadvantages qualified applicants, removes it, adjusts the process, and adds human review and appeal routes.

Not every difference in outcomes proves unlawful discrimination, and no single fairness metric works for every context. Some fairness criteria conflict mathematically, so acceptable trade-offs require explicit governance and domain expertise.

Bias can arise from historical conditions, problem framing, sampling, labels, measurement, proxies, model design, optimisation objectives, deployment context, or human interpretation. It is socio-technical: even statistically accurate models can create harmful outcomes when categories, thresholds, or uses are inappropriate.