Deployment bias is a systematic disparity that occurs when an artificial intelligence system serves some populations better than others. It is one possible source of bias across the AI pipeline, alongside data collection, labeling, and model development. Bias reflects a mismatch between model outputs and the diverse populations those outputs represent; it can produce different results across attributes such as race, gender, class, age, or religion. section 3.2
A deployment-stage disparity
A system that performs better for one population than another illustrates deployment bias. This differs from labeling bias, which arises when annotators’ assumptions affect the labels used to train a model. section 3.2
