AI Tools and Sentencing Decisions: Analysis of Pennsylvania Sentencing Data for Legal Financial Obligations

• In Event: AI, Equity, Ethics, and Inclusion

Abstract

As financial penalties can impose disproportionate burdens, especially on economically vulnerable populations, understanding the factors influencing judicial decisions is critical. At the same time, the growing interest in artificial intelligence (AI) across domains has prompted increased exploration of its applications in the criminal justice system, including legal decision-making and sentencing determinations. Standards and guidelines for imposing legal financial obligations (LFOs), including monetary fines and restitution, vary widely across jurisdictions, with some states offering little to no formal direction. Some states, Pennsylvania in particular, have established court precedents requiring judges to consider specific factors related to a defendant’s ability to pay when determining LFO amounts. This study specifically examines the presence of bias in judicial determinations of LFOs, within criminal sentencing in the state of Pennsylvania, with a particular focus on defendants’ ability to pay. Data from the Pennsylvania Commission on Sentencing from 2001 through 2012 (N=702,755) was used, focusing on individuals with LFOs imposed as their only sentence (N=303,662). Results from our previous theory-driven work indicated that, while imposition of LFOs appeared racially-neutral, there were racially-disparate impacts for individuals’ ability to pay. Additionally, the LFOs imposed did not uniformly follow recommendations for lower and upper limit recommendations. In the present study, we evaluated whether AI-based predictive models to assess whether these specific judicial sentencing decisions align with expected LFOs to be imposed based on state legal precedents for factors to consider when determining these amounts. We additionally trained and tested multiple machine learning models, including feature selection, to estimate appropriate LFO amounts based on defendants’ financial capacity, and these predictions were then compared against actual fines imposed by judges. Lastly, we employed AI-driven interpretability techniques to identify patterns and potential sources of bias within judicial sentencing. By analyzing discrepancies between model predictions and judicial outcomes, we highlight areas where extralegal or disproportionately weighted factors may influence LFO determinations in violation of the required guidelines in the state of Pennsylvania. Our findings contribute to the broader discourse on fairness and accountability in sentencing, while also assessing the potential role of AI as a tool for detecting and mitigating bias in legal systems.

Author

• Divya Ramjee
  Presenting Author