Data bias and public sector AI: a diagnostic simulation of performance and equity trade-offs in predictive policing

Purpose Public sector AI adoption is uneven and difficult to sustain beyond experimentation, in part due to the persistent problem of biased and low-quality data. These biases often stem from historical inequities embedded in datasets. This study examines whether a data-centric AI approach, using data augmentation via synthetic oversampling, can reveal how training-data imbalance shapes performance and equity trade-offs within predictive policing systems. Design/methodology/approach We combine arrest, demographic and mapping data for Chicago. The arrest data from the Chicago Police Department are highly imbalanced. Using spatial regression models and the Synthetic Minority Oversampling Technique (SMOTE), we construct diagnostic simulations to evaluate model performance and spatial concentration under different scenarios. Findings Our models show that predictions trained on imbalanced arrest data can increase spatial co-concentration between predicted enforcement activity and the Black residential distribution. Applying SMOTE reduces the activity ratio, our exploratory measure of this co-concentration, by up to 15%, albeit with a moderate performance trade-off. Practical implications For public agencies, incorporating data-centric AI strategies can help identify how model outputs shift when underlying data distributions are adjusted. While not a substitute for structural reform or governance oversight, these approaches offer a concrete step for practitioners tasked with implementing AI systems. Originality/value Utilizing synthetic oversampling, we demonstrate how data-centric methods can examine class imbalance and data bias, critical yet understudied barriers to responsible AI implementation in government.