Strengthening Artificial Intelligence Governance through Ethical Handling of Sensitive Data: An Applied Study on Text Classification and Differential Privacy

Abstract

This research develops a comprehensive hybrid framework to enhance Artificial Intelligence governance by ethically managing sensitive textual data through advanced classification techniques. Focusing on natural language processing (NLP) applications, the study integrates rule-based systems, logistic regression, and transformer-based models, notably BERT, to address the challenges of identifying and handling sensitive information within complex and ambiguous linguistic contexts. Experimental results demonstrate that the hybrid model attains an overall classification accuracy of 91%, with precision and recall scores of 89% and 94%, respectively, achieving an F1-score of 92%. These metrics reflect the model’s robustness in real-world scenarios where explicit textual indicators are often lacking. Individually, the rule-based approach excels in precision (98.6%) for clearly identifiable sensitive content, logistic regression ensures perfect recall (100%), capturing all sensitive instances albeit with increased false positives, and the BERT model achieves perfect precision, effectively minimizing false alarms. The hybrid approach synergizes these strengths, resulting in a balanced and reliable classification system. The study further explores the integration of differential privacy via a differentially private logistic regression model using the diffprivlib library, assessing privacy-utility trade-offs at varying privacy budgets (ε = 3, 5, 6). Results reveal that stronger privacy guarantees (lower ε) reduce classification accuracy (78% at ε=3), while looser privacy constraints (ε=6) approach non-private model performance (97% accuracy). These findings underscore the potential of combining hybrid NLP models with differential privacy to deliver scalable, trustworthy, and privacy-preserving AI systems. The proposed framework holds significant relevance for sensitive domains such as healthcare, public administration, and corporate governance, where balancing data privacy and AI performance is critical. Future research should extend these findings by exploring additional privacy configurations and validating the approach against diverse real-world datasets to optimize the equilibrium between privacy protection and analytical effectiveness.