Why Gender Bias Persists in Machine Learning Models

This study examines how gender bias persists in podcast recommendation models, even when gender is excluded as a feature. Using latent space visualizations, statistical tests, and multiple bias metrics (EAA, GEAA, DEAA, R-RIPA), the researchers found that gender associations remain encoded in embeddings, influencing predictions and classifications. The findings highlight that simple mitigation—like removing gender data—doesn’t erase bias, emphasizing the need for multi-directional bias testing and more nuanced fairness interventions in AI systems.