AI Literacy / Glossary

Algorithmic bias

Algorithmic bias is when an AI system or other automated tool gives results that are unfairly better or worse for some groups of people, often because of skewed or missing examples in its data.

AI learns from past data. If that data under-represents some people, or reflects unfair decisions from the past, the model can learn and repeat those patterns, for example working less well for some skin tones, accents or names.

Bias can also come from how a problem is set up: what the system is asked to predict, which information it uses and who checks the results. It is usually not intended, which is why it needs to be tested for.

Fair AI needs varied data, testing on different groups and human review of important decisions. In many places, people also have rights to ask how an automated decision about them was made.