Why can AI be biased?
Models can repeat inequalities found in their data, their goals and the way they are used.
What is it?
An AI system is biased when it works worse for some groups of people or treats them unequally. Bias comes from the data, the chosen goals and the way the system is used. No system is entirely unbiased, but bias can be measured and reduced.
An everyday example
A face recognition system misidentifies people with darker skin more often because there were fewer such photos in its training data.
Why does it matter?
In hiring, lending, policing or health care, bias can deprive people of opportunities or rights.
A common misconception
That high average accuracy means a fair system. The average can hide poor results for a particular group.
In more detail
Where does bias come from?
From how data is chosen, missing examples, labelling, the measures of success and the social context.
Is it possible to build completely unbiased AI?
There is no simple, universal measure of fairness. Different goals can pull against each other.
Where is it especially dangerous?
In decisions about jobs, loans, education, surveillance or public services.
What can we do?
Test the system on different groups, track its errors, involve the people affected and set up a way to complain.
Remember: Good average accuracy can hide poor results for a particular group.
Reliable sources
Last reviewed 1 November 2026. An explainer, not professional, legal or medical advice. For rules and figures that change, check the official version.