Implicit Bias
Unconscious or unintentional biases embedded in training data or model design that can lead to discriminatory outcomes.
Definition
Biases introduced by societal, cultural, or sampling factors that data-curation processes do not actively surface. Implicit bias may lurk in labeler judgments or historical records. Governance calls for blind-labeling protocols, diverse annotation teams, and regular bias-detection scans to uncover and correct these hidden drivers of unfair outcomes.
Real-World Example
A sentiment-analysis model trained on social-media posts reflects an implicit bias: posts from certain dialects are labeled more negatively. The team institutes blind labeling (removing author metadata) and recruits diverse annotators, reducing misclassification rates for dialectal text by 40%.
.svg)
