# What is Bias?

Also called Algorithmic Bias, AI Bias, Model Bias.

Bias in AI refers to systematic differences in a system's behavior or accuracy across groups, contexts, or inputs, particularly where those differences are unjustified and cause harm. It originates in training data, labeling choices, objective design, and deployment context rather than in any single component, and it is measured against a chosen fairness definition.

The word carries two distinct meanings that get conflated. Statistical bias is a systematic deviation between an estimate and the underlying value, a neutral technical property. Social bias is a pattern of behavior that disadvantages a group, a normative judgment. A system can exhibit either without the other, and conversations go wrong quickly when the two senses are mixed.

Sources are cumulative across the pipeline. Training corpora reflect who produced the text and in which languages. Human labelers bring their own assumptions. Objective functions optimize aggregate performance, which can be achieved while performing poorly for a small subgroup. Deployment adds its own effects when a system trained on one population is applied to another, or when its outputs feed back into future training data.

Measurement forces a choice that cannot be avoided. Several formal fairness criteria, such as equal error rates across groups and equal predictive value across groups, have been proved mathematically incompatible except in special cases. There is no configuration that satisfies all of them at once, so any team measuring fairness is implicitly selecting which definition matters for their context, and stating that choice openly is better than leaving it implicit.

The field is actively contested. Disagreement covers which harms count, whether removing group attributes helps or hides the problem given correlated proxies, whether measured differences reflect the model or the world it was trained on, and how much responsibility sits with model providers rather than deployers. Vendors sometimes claim unbiased systems, a claim that is not meaningful without naming a definition and showing measurements.

## Key points

- Statistical bias and social bias are different claims
- Sources accumulate: data, labels, objective, deployment context
- Common fairness criteria are provably incompatible with each other
- Removing an attribute does not remove correlated proxies
- An unqualified claim of no bias is not a meaningful statement

## In practice

A resume screening model is evaluated overall and performs well. Broken out by applicant group, its false negative rate is noticeably higher for one group whose resumes more often use non-Western institution names underrepresented in training data. Removing the names does not close the gap, because institution, phrasing, and formatting carry the same signal. The measured difference only became visible once results were disaggregated.

## Related terms

- [Model Evaluation](/en/glossary/model-evaluation)
- [Alignment](/en/glossary/alignment)
- [Content Moderation](/en/glossary/content-moderation)
- [AI Governance](/en/glossary/ai-governance)
- [Explainability](/en/glossary/explainability)

[Back to the AI Glossary](/en/glossary)