Entry 14 of 44
Bias
biais — oblique, slant; origin uncertain (possibly Provençal)
Old French
From Old French biais: oblique, slanted. Root concept: a systematic slant or departure from the straight — an inclination that produces skewed outcomes.
Established meaning
In statistics: systematic error in a method producing results that deviate from truth in a consistent direction. In law and civil rights: prejudice that produces discriminatory outcomes. Both uses imply a harm to identifiable parties.
AI industry deployment
Statistical disparity in model outputs across demographic categories. 'Bias mitigation' often means adjusting outputs to pass disparity metrics. The civil rights dimension — what is owed to persons who experience discriminatory outcomes — is frequently absent from the technical treatment.
Semantic gap
The person harmed
Statistical debiasing that achieves parity on benchmarks while failing to address real-world discriminatory impact treats a metric as the thing the metric is supposed to measure.