Algorithmic bias causes documented harm: 45% lower loan approval rates for minorities (CFPB 2024), CVs rejected based on gender. The EU AI Act classifies scoring and rating systems as high-risk.
Model learns discriminatory patterns present in training data. Example: credit models trained on past decisions reflecting systemic discrimination.
Under-representation of groups in data. Example: facial recognition datasets with 80%+ white male faces → 34% error on dark-skinned women (MIT Media Lab, Buolamwini & Gebru).
Discriminatory proxy: using zip code as a credit risk variable indirectly captures ethnicity. Statistical correlation is not legal causality.
Model influences future data. Example: police recommendation system → more patrols in certain neighbourhoods → more arrests → more data "validating" the bias.
Open-source Python library for measuring and reducing ML classifier inequities. Metrics: demographic parity, equalized odds, equal opportunity.
Open-source, production-readySuite of 70+ fairness metrics and 10+ bias reduction algorithms. Pre-processing, in-training, post-processing approaches.
70+ fairness metricsLocal and global decision explainability. Required for EU AI Act high-risk systems (Art.13: transparency, Art.14: human oversight).
Required by EU AI Act Art.13Molderez Consult SRL audits and secures your AI compliance (EU AI Act, GDPR).
Compliance audit