AI model bias appears when a model consistently favors, omits or misrepresents particular groups, viewpoints or outcomes. It can arise from training data, labeling, optimization goals, evaluation choices and how an application frames the user's options.
Bias cannot be removed through one universal metric because fairness depends on context and affected people. Responsible deployment uses representative evaluation, source transparency, stakeholder review, monitoring and a process for contesting decisions or correcting harmful patterns.
ELI5
Model bias is a repeated tendency in an AI model's results that unfairly favors, overlooks or misrepresents certain people or outcomes. It can come from the training data, labels, goals, evaluation choices or the way a product uses the model.
For example, a hiring model trained mostly on past employees from one group may rank similar applicants more highly. Teams should test representative cases, involve affected people and provide a way to challenge or correct harmful decisions.
Where does AI model bias come from?
It can come from historical data, sampling gaps, labels, model objectives, product design, user prompts and the social context in which outputs are used.
How can organizations test for model bias?
They can evaluate representative cases across affected groups, compare outcomes, inspect failure patterns and include domain experts and impacted communities in review.
