How AI Systems Inherit Human Bias Through Flawed Historical Training Data

Machine learning models can reproduce societal biases not through explicit programming but by learning patterns embedded in historical training data. A loan-scoring AI, for example, may assign lower credit scores to women simply because past data reflected discriminatory lending practices, even if gender is never a direct input. This phenomenon, known as proxy discrimination, occurs when seemingly neutral variables like working hours or postcode are statistically correlated with protected characteristics such as gender or race. Removing sensitive attributes from datasets does not eliminate bias, as models can reconstruct those patterns through correlated data points. Researchers are developing algorithmic fairness techniques, including pre-processing methods that rebalance training data, to address these deeply embedded disparities.