Algorithmic Bias
Systematic errors in a model that skew outputs in ways that are wrong or unfair, often rooted in training data, feature selection, or model design.
Algorithmic bias is a systematic error in an AI model that causes it to produce skewed, inaccurate, or unfair outputs. Bias can enter from the training data, from the choice of input variables, from how the model is trained, or from how its outputs are used in a business process.
Bias is related to but distinct from discrimination. Bias is an input-side or model-side problem: the model is systematically wrong. Discrimination is an outcome-side problem: the model produces different results for protected classes. A biased model can be discriminatory, and a discriminatory model is usually biased, but the fixes can differ.
Testing for algorithmic bias typically involves checking accuracy across groups, measuring disparate impact, and examining whether the model’s errors fall more heavily on some populations than others. See our glossary entries on algorithmic discrimination and bias testing.