Cultural Bias in Machine Learning Systems: A Philosophical and Empirical Study of Algorithmic Knowledge Production

Abstract

Machine learning systems are increasingly functioning as epistemic infrastructures in high-stakes domains such as criminal justice, healthcare, finance, and employment. Despite this, their outputs are frequently treated as objective and neutral forms of knowledge. This study advances a synthesis of empirical and philosophical inquiry into cultural bias in machine learning, arguing that algorithms operate as sociotechnical agents embedded within historically situated structures of power and representation. Using the COMPAS Recidivism dataset (N = 7,214), a quantitative experimental design was employed to examine predictive disparities across protected attributes, specifically race and sex. Logistic Regression and Random Forest models were implemented within a controlled preprocessing pipeline and evaluated using standard performance metrics (accuracy, precision, recall, and F1-score), alongside subgroup fairness measures including false positive rates (FPR), false negative rates (FNR), and disparate impact ratios. To ensure robustness, subgroup disparities were further assessed using statistical significance testing. While overall model performance was moderate in aggregate metrics, subgroup analysis revealed consistent and structured disparities: African-American defendants exhibited elevated false positive rates, whereas females and underrepresented racial groups experienced disproportionately high false negative rates. These patterns persisted across model architectures, indicating that bias is structurally embedded in the data rather than solely a function of model design. However, extreme subgroup values should be interpreted with caution due to potential sample size imbalances within certain demographic categories. The findings challenge the assumption of epistemic neutrality in algorithmic systems, demonstrating that machine learning models participate in the cultural production of knowledge by reproducing historically grounded classifications and power asymmetries. The study argues that algorithmic outputs should be evaluated not only in terms of predictive performance but also through fairness-aware and context-sensitive frameworks that account for their broader ethical and epistemological implications.