Regulating Artificial Intelligence in the Age of Autonomous Decision-Making: A Legal and Ethical Framework

Abstract

The growing use of autonomous systems that make decisions in high-stakes institutional contexts has raised growing concerns about predictive fairness, accountability and regulatory oversight. While it is often true that machine learning models improve classification accuracy, the distributive consequences of these models are controversial. In this study, the empirical performance and fairness implications of supervised learning models in the context of judicial recidivism risk assessment are empirically evaluated using a publicly available publicly available COMPAS recidivism dataset distributed via Kaggle. Linear and non-linear classifiers (Logistic Regression, Random Forest, Gradient Boosting, and Support Vector Machine with RBF kernel) were benchmarked with respect to accuracy, ROC-AUC, Cross-Validation, Calibration Analysis and Subgroup Fairness Diagnostics. The Random Forest model showed a better discriminative power (AUC = 0.9065) and stable cross-validated generalization performance; however, the differences in the false positive rates of the different demographic groups were not negligible in spite of high performance in terms of accuracy aggregate. Statistical validation was done to ensure that difference in performance was systematic and not stochastic. These results show how there is a structural tension between predictive optimization and distributive equity, which point to the fact that technical improvements in performance cannot address questions of fairness. The study is consistent with risk-based approaches to regulation that integrate techniques of statistical validation, fairness audits and institutional accountability mechanisms. Effective management of autonomous AI systems therefore requires a balancing act of the predictiveness of the AI system and an equitable impact to foster legitimacy in high stakes decision environments.