Bias is systematic skew in an AI system's answers – for example with regard to gender, age, ethnicity or language. It arises because the system has learned from data that is itself skewed, and it does not disappear by itself.
Bias is not an error in the classic sense, but a condition: models learn from the texts and images of the past, and the past is not neutral. A recruitment model can favour certain profiles, a language model can describe professions in stereotypical terms, and an image tool can turn 'a leader' into a particular kind of person.
So the crucial thing is not to find a system without bias – no such system exists – but to know where the skews can do harm, and to put human oversight there. The greater the consequences of a decision for people, the less it should be left to the machine alone.
In practice
In practice, bias is handled with three measures: critical questions to the vendor, spot checks of the system's answers in your own workflows – and clear rules for which decisions always require a human.
How to explain it to management
»The AI's answer is a mirror of its data. Our responsibility is to know when the mirror is distorted – and act accordingly.«
Bias is one of the reasons behind the requirements in the EU AI Act – and a recurring theme in my work on adoption in the intelligent organisation.