Principles of Data Governance for Ethical AI Systems

This thesis proposes a synthetic data governance framework for ethical AI systems that connects data quality, bias mitigation, privacy protection, transparency, and accountability. The theoretical part systematizes the core concepts and comparatively analyzes four frameworks and standards (DAMA-DMBOK, GDPR, NIST AI RMF, and the EU AI Act). On that basis, the five-pillar AI-DG Framework is formulated. The practical part uses a repeated additive intervention design with five conditions on the UCI Adult Census Income dataset across 30 repetitions (seeds 42–71) using logistic regression and random forest. The combined pipeline reached k-anonymity k = 5 and about a 97.5% reduction in quasi-identifier combinations on the training data, while reweighing produced the strongest improvement in sex fairness for logistic regression (disparate impact rose from a baseline of 0.282 to 0.608, yet remained below the 0.8 threshold). However, the combined pipeline did not achieve the best predictive performance and should be interpreted as one compromise configuration under the chosen fairness and privacy metrics rather than as a universally optimal solution. The trade-off narrative holds selectively: reweighing and the combined pipeline slightly reduced performance, whereas the privacy-only condition actually improved it with statistical significance. The findings therefore serve as an illustration of selected intervention trade-offs on one public dataset under a custom repeated-holdout protocol, not as a general validation of the whole framework. The main contribution is the design of an integrative AI-DG Framework and its transparent empirical confrontation, demonstrating both the potential and the limits of data-layer interventions.