Glossary
Algorithmic Bias
Systematic skew in an algorithmic system’s outputs that can produce unjustified disparate effects across people or groups.
Last reviewed 2026-08-22
In plain language
Bias can arise from training data, problem formulation, proxy variables, feedback loops, or unequal error rates. Not every statistical disparity is unlawful discrimination, and not every fair-looking metric is substantively just. Debiasing techniques are context-specific and can trade off different notions of fairness. Human-rights and equality law, not only technical metrics, remain the reference for public-sector and high-stakes uses. A commonly cited pattern illustrates several of these mechanisms at once: a hiring-screening model trained on a company's historical resume data can learn to reproduce that company's past hiring patterns, including any prior underrepresentation of a protected group, even when the protected characteristic itself is never given to the model, because correlated proxy variables—a university's demographics, a gap in employment history, even certain phrasing—can carry the same signal indirectly. Removing the obvious variable does not remove the bias; it can make it harder to detect. This is why algorithmic bias and algorithmic accountability are related but distinct terms: bias describes a property of a system's outputs, while accountability describes the institutional capacity to identify who is responsible for that property, investigate it, and provide a remedy to someone affected by it. A system can be measurably biased with no accountability mechanism attached, and a system can have a formal accountability process that never actually tests for bias—governance work needs to ask about both separately.
Why it matters for AI governance
Bias is one of the most documented AI harms in employment, credit, policing and public services. Governance responses include testing, documentation, prohibited practices and equality-impact duties.