How Bias Creeps In
Training data reflects historical conversion. If historical conversions favored certain demographics, the model learns that preference. New leads from under-represented groups score lower — reinforcing the bias.
Detection
Audit score distributions across protected characteristics (where legally appropriate to collect — US anti-discrimination framework varies). Disparate impact analysis. Significant gaps signal bias.
Mitigation
Remove protected characteristics from training features. Audit for proxy variables (zip code correlating with race). Retrain with balanced data. Consider fairness-aware ML techniques.
EU AI Act Intersection
Lead scoring affecting access to essential services is Annex III high-risk. Conformity assessment includes bias audit. US orgs with EU-user data must comply. Start the audit before August 2.