Definition · Plain-language
AI bias
AI bias is a systematic, unfair skew in an AI system’s outputs that arises from its data, design or deployment.
The step most authors miss
Doing CRediT right? Don’t stop at the statement.
A CRediT statement credits you inside one paper. The recognition CRediT was built for happens when those roles are tied to you, persistently. Sign in with your ORCID — free — and claim your CRediT contributions on casrai.org, the home of the standard. They become a verified, portable part of your identity, not a line that disappears into one PDF.
Free: claim your contributions, then export a journal-ready CRediT statement, schema.org structured data, JATS XML, CSV or BibTeX — and preview your public profile. A membership publishes that profile publicly and verifies the journals you serve.
Where bias comes from
Bias can enter at several points. Data and sampling bias arises when training data under-represents or misrepresents groups, or encodes historical inequities, so the model learns a skewed picture of the world. Algorithmic bias arises from design choices — features selected, objectives optimised, thresholds set — that systematically favour some outcomes. Human or label bias enters through the judgements of those who annotate data or define the problem. Deployment bias appears when a system is used in a context, or on a population, different from the one it was built and validated for.
Why it matters
Because AI systems operate at scale and often opaquely, biased outputs can entrench unfairness across many decisions before anyone notices. The harm is not only ethical but practical: skewed systems erode trust, can breach equality and non-discrimination expectations, and degrade real-world performance for affected groups. Standards bodies treat bias as a core trustworthiness concern — the NIST AI RMF lists managing harmful bias among its trustworthiness characteristics, and NIST has published dedicated guidance distinguishing statistical, systemic and human bias in AI. Managing it is therefore central to responsible AI rather than optional.
Mitigation across the lifecycle
No single fix removes bias; mitigation is layered across the lifecycle. At the data stage it means seeking representative, well-documented data and examining it for skew. At design and training it means testing models for disparate performance across groups and weighing fairness alongside accuracy. Before and after deployment it means independent assessment, clear documentation, human oversight of high-impact decisions, and monitoring for drift that reintroduces bias over time. Crucially, mitigation requires defining what fairness means for the specific context, since fairness measures can conflict and no system is fair in every sense at once.
Key facts
At a glance
- Definition: systematic, unfair skew in AI outputs from data, design or deployment
- Data/sampling bias: unrepresentative or skewed training data
- Algorithmic bias: skew introduced by model design choices
- Human/label bias: bias from annotators or problem framing
- Deployment bias: use in a context different from validation
- Mitigation: representative data, fairness testing, oversight, monitoring
Common misconceptions
What people often get wrong
Often heard: AI is objective, so it cannot be biased.
Actually: AI systems learn from data and design choices made by people, both of which can encode skew. Far from removing bias, AI can amplify it at scale, which is why managing harmful bias is a recognised trustworthiness concern.
Often heard: Removing sensitive attributes like race or gender makes a model unbiased.
Actually: Models can infer sensitive attributes from correlated features, so simply dropping them does not guarantee fairness. Effective mitigation requires testing outcomes across groups, not just hiding inputs.
Often heard: A model can be made perfectly fair on every measure.
Actually: Different fairness measures can be mathematically incompatible, so trade-offs are unavoidable. Fairness must be defined for the specific context rather than assumed to be achievable in every sense simultaneously.
Going deeper
