Bias in AI refers to systematic unfairness in a model’s outputs, usually caused by patterns in its training data that reflect real-world prejudices or imbalances. (In a separate technical sense, “bias” also names a basic parameter inside models.)
What it means in plain English
AI models learn from data, and if that data reflects historical or societal biases, the model can absorb and reproduce them — often while appearing neutral and objective. Because the output looks like impartial computation, this bias can be harder to spot and challenge than an obviously prejudiced human decision.
It is one of the most serious ethical challenges in AI, especially when models are used in decisions about people.
A simple example
If a hiring model is trained on a company’s past hiring data, and that history favoured certain groups, the model can learn to favour them too — presenting a biased outcome as a neutral, data-driven recommendation.
Why it matters
Understanding bias is essential for using AI responsibly. It is why fairness must be actively tested for rather than assumed, why human oversight matters for decisions affecting people, and why “the algorithm decided” is never an acceptable justification on its own.
Related terms
- Training Data — the usual source of bias.
- Hallucination — another key limitation of AI.
- Model — what can encode and reproduce bias.
Frequently asked questions
Where does AI bias come from?
Most often from the training data: if the data reflects historical or societal biases, the model can learn and reproduce them. Bias can also arise from how a problem is framed or how data is labelled.
Why does AI bias matter?
Biased models can produce unfair or discriminatory outcomes — in hiring, lending, or policing, for example — which is why detecting and mitigating bias is a central concern in responsible AI.
