Bias (in AI)
When an AI system gives results that are unfair or skewed, for example working better for some groups of people than others, often because of patterns in its training data.
AI systems learn from examples, and examples come from the real world, with all its gaps and unfairness. If a system mostly saw one kind of person, place or situation during training, it tends to do worse on everything else. Choices made by the people who design it can add bias too.
Well-documented cases include face recognition tools that made more mistakes on darker-skinned faces, and a hiring tool that learned to favour men because it was trained on a company's past hiring decisions. Image generators asked for "a doctor" or "a nurse" have also leaned on stereotypes.
Bias is not only about obvious prejudice. It can be subtle and hard to notice from the outside. A computer producing a result does not make it objective. When AI is used for decisions about people, such as jobs, loans or benefits, it is fair to ask how it was tested and who checks it.
Related: Training data, Algorithm, Machine learning, AI Act (EU)