An Original Contribution to Explainable AI Methods for Predictive Policing Systems in the Nigerian Security Context

Predictive policing AI systems are beginning to be considered by Nigerian security agencies as tools for resource allocation and crime prevention, but their deployment raises serious accountability, bias, and rights concerns that require explainability mechanisms designed for the specific Nigerian security governance context. This study develops original explainable AI methods for predictive policing systems adapted to the Nigerian security context. Four original contributions are presented: a geographic bias detection framework that identifies racial and ethnic proxy variables in crime prediction models trained on Nigerian arrest records, correcting for systemic over-policing patterns in historically marginalised communities; a contrastive explanation generation method that produces actionable counterfactual explanations for area risk scores in terms comprehensible to non-technical police commanders; a participatory explanation design methodology that engages community representatives in co-designing explanation formats, tested through workshops in six communities across Plateau, Lagos, and Kano States; and a formal accountability traceability architecture linking every prediction to its feature attributions, training data sources, and human decision override records. The framework was applied to a prototype predictive policing model trained on five years of crime reports from two state police commands. Bias evaluation revealed that ethnically diverse neighbourhoods were 2.7 times more likely to receive high-risk scores under the baseline model after controlling for actual crime rates. The debiased model reduced this disparity to 1.3 times while maintaining prediction accuracy. The study provides original XAI methodology applicable to high-stakes security AI contexts globally.

Keywords: explainable AI, predictive policing, algorithmic bias, Nigeria, AI accountability