Examples
Worked examples
- Is an instance
A model tested against a held-out, group-labelled evaluation set showing a measurably higher error rate for one demographic group than another
- Is an instance
A language model trained predominantly on English-language, high-resource-region text showing systematically worse performance on lower-resource languages
Counter-examples
Looks similar, but isn't
- Not an instance
A claim that a system is 'fair' or 'unfair' with no accompanying measurement of a specific output disparity is a fairness claim, not a bias measurement
- Not an instance
A single-group accuracy figure with no cross-group comparison does not, by itself, demonstrate bias -- bias is inherently a comparative, cross-group or cross-context measurement
Editorial commentary
AI bias, in the sense used across responsible-AI frameworks, is an empirical property: does a system’s output differ systematically and unjustifiably across groups, tasks, or contexts, and if so, by how much and in which direction. It is measurable, at least in principle, given the right evaluation data — which is what distinguishes it from AI fairness, a separate, contested normative question about which measured disparities should be considered acceptable.
The NIST AI RMF’s three categories
The NIST AI Risk Management Framework (AI RMF 1.0, published January 2023) identifies three categories of bias relevant to managing AI risk, and is explicit that bias is not confined to demographic representativeness alone. Systemic bias can appear in datasets, in an organisation’s practices across the AI lifecycle, or in the broader societal context an AI system is deployed into — it precedes and is broader than any single model. Computational and statistical bias arises from technical choices in how a model is trained and evaluated — sampling, feature selection, optimisation objectives, and evaluation-set design can all introduce skew independent of any societal factor. Human-cognitive bias enters through the people who design, build, label data for, and deploy an AI system, reflecting how individuals perceive and interpret information and make decisions during that process. NIST notes explicitly that biased outputs can arise without any prejudicial or discriminatory intent on anyone’s part, which is why bias needs active, structured management rather than an assumption that good intentions are sufficient.
Where composition feeds into it
Training data composition is one of the most direct upstream drivers of computational/statistical bias — a corpus skewed in language, geography, or demographic representation tends to produce correspondingly skewed model behaviour, though bias can also arise purely from algorithmic or evaluation design choices even with well-balanced data.
How this differs from AI fairness
Bias is what you measure; fairness is the contested judgment about what to do with the measurement. A system can have a precisely documented bias (a specific, quantified error-rate gap between two groups) while reasonable people, applying different but individually defensible formal fairness criteria, disagree about whether that gap makes the system unacceptable for a given use. Treat any bias measurement as an input to a fairness discussion, not a substitute for one.
Also known as
Algorithmic bias · Model bias
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="AI bias"
vocab-term-identifier="https://casrai.org/dictionary/term/ai-bias" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/ai-bias",
"name": "AI bias",
"identifier": "https://casrai.org/dictionary/term/ai-bias",
"description": "A measurable, descriptive property of an AI system's outputs: a systematic skew that produces unjustified differences in accuracy, treatment, or representation across groups, tasks, or contexts, arising from training data composition, algorithmic design choices, or deployment context -- distinct from ai-fairness, a contested normative judgment about which disparities are acceptable.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/genai-disclosure#set",
"url": "https://casrai.org/dictionary/term/ai-bias",
"alternateName": [
"Algorithmic bias",
"Model bias"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T01:55:38",
"dateModified": "2026-08-22T14:52:08",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
