Ethics and Fairness in Machine Learning: Equity Guidelines for Brazilian Public Policies.

Abstract

The application of machine learning technologies in public policy formulation represents a significant advance in contemporary government management, but raises critical issues related to equity, transparency, and algorithmic justice. The present study investigates the presence of systemic biases in data and predictive models applied in Brazilian public policy scenarios through a comprehensive audit using open government data. A pipeline-structured methodology was developed that covers ten social scenarios in critical areas such as education, health, work, and technology, using data from five Brazilian government agencies (IBGE, MS, MEC, INEP, CETIC.BR). Eight supervised binary classification algorithms were evaluated. To rigorously address the severe class imbalance inherent in public policy microdata, the final predictive models were selected strictly based on the highest 5-fold Cross-Validated Area Under the ROC Curve (CV AUC-ROC) rather than global accuracy. This unified criterion resulted in the selection of Gradient Boosting for seven social scenarios and Random Forest for three scenarios (Food Insecurity, Digital Inclusion, and Tech Vulnerability) as the most robust models, which were subsequently subjected to systematic fairness auditing. The results indicate the presence of consistent patterns of disparity among sensitive demographic groups, with frequent violations of established disparity limits, particularly in the Demographic Parity and Predictive Equality metrics. Because the reference group is the modal category of each sensitive attribute ---Brown individuals in every racial audit---the reported ratios express deviation from the predictive pattern of that group, and such deviations affect all contrasted categories: minority groups (black, yellow, and indigenous people) record the most extreme cases, including predictive collapses, while the white group registers some of the largest deviations in positive prediction and false positive rates. These patterns show concrete risks of algorithmic discrimination and perpetuation of historical inequalities. The Equal Opportunity metric suggests better equity for the gender sensitive attribute in eight of the ten scenarios analyzed, the exceptions being Poverty Prediction (CENSO) and Live Births Vulnerability (SINASC), where the metric fails against the sex reference of the respective dataset. This research contributes significantly to the field of ethical and responsible Artificial Intelligence in the Brazilian context, providing empirical evidence of the presence of systematic biases in government data and models and establishing a replicable protocol for fairness audit in automated decision support systems. The findings demonstrate the urgent need for the implementation of continuous audit practices in the life cycle of Machine Learning projects in the public sector. Furthermore, this study identifies a critical gap for future research and policy-making: the urgent necessity to formulate specific legal instruments and strengthen existing data protection frameworks (such as the LGPD in Brazil) to formally enforce algorithmic equity and regulate automated decision-making. By addressing this regulatory void, we propose guidelines for the development of more just and equitable Artificial Intelligence solutions that are rigorously aligned with democratic principles, legal accountability, and fundamental rights.