Impact of AI-assisted sentencing on judicial fairness and bias in appellate courts

Thesis Abstract

The study investigates how AI-assisted sentencing tools influence judicial fairness and perceived bias within appellate courts, addressing growing concerns that algorithmic recommendations may entrench or mitigate existing disparities in sentencing outcomes. The aim is to assess whether integrating AI-aided guidance into appellate decision-making affects consistency, transparency, and equity across diverse defendant profiles, while identifying mechanisms that contribute to potential bias, either amplifying or counteracting human judgment. Specific objectives include (1) evaluating the relationship between AI-assisted sentencing recommendations and appellate outcome variance across case types (violent vs. non-violent offenses) and offender characteristics (demographics, prior records); (2) examining appellate judges’ reliance on AI outputs through qualitative interviews and observation of deliberation sessions; (3) testing for changes in sentencing alignment with statutory mandates and sentencing guidelines; and (4) identifying governance, governance, and procedural factors that influence fairness in AI-enabled appellate processes. The study adopts a mixed-methods design, combining quantitative analysis of appellate decisions with qualitative insights from key stakeholders. The population comprises appellate court cases from five major jurisdictions over a five-year period where AI-assisted risk assessments or sentencing recommendations were available to judges, yielding a dataset of approximately 2,400 cases and accessible de-identified judge deliberation transcripts for 60 appellate judges. Quantitative data are sourced from official court records, AI tool outputs, and sentencing outcomes, while qualitative data are collected through semi-structured interviews with appellate judges, clerks, prosecutors, defense attorneys, and AI system developers. Data collection instruments include a coding framework for AI utilization, a variables matrix capturing offender and case characteristics, and interview protocols designed to elicit perceptions of fairness, transparency, and trust in AI outputs. Validity and reliability are addressed through triangulation of court-record data, cross-validation of AI outputs with independent risk assessment benchmarks, and pilot testing of instruments with expert panels. Data analysis employs multilevel regression models to account for clustering within jurisdictions and case heterogeneity, including hierarchical linear modeling to examine the impact of AI reliance on sentencing variance, while controlling for case severity and demographic covariates. Additional techniques include propensity score matching to compare AI-assisted with non-AI-assisted appellate decisions where feasible, and sensitivity analyses to assess robustness to alternative model specifications. The qualitative component uses thematic analysis of interview transcripts and deliberation notes, guided by the procedural justice and algorithmic accountability theoretical frameworks, with intercoder reliability checks and member-checking to ensure credibility. The theoretical backbone draws on theories of procedural justice, algorithmic fairness, and bias in decision-making, including Rawlsian concepts of fairness and the fairness through awareness approach, alongside the socio-technical systems perspective to illuminate how organizational practices shape AI effects. Expected findings anticipate a nuanced impact AI-assisted recommendations may reduce sentencing variance in some jurisdictions, enhancing consistency, but may also interact with human biases to disproportionately influence outcomes for certain demographic groups or offense types, particularly where data inputs reflect historical inequities. The study contributes to knowledge by elucidating the conditions under which AI tools support or hinder judicial fairness in appellate contexts, offering empirical benchmarks for measuring algorithmic accountability, and informing policy design for governance, oversight, and transparency of AI-assisted sentencing. Practical implications include recommendations for standardized disclosure of AI inputs in appellate deliberations, governance structures to monitor bias, and training programs to enhance judges’ critical engagement with algorithmic advice. The conclusion emphasizes that while AI-assisted sentencing has the potential to improve consistency and objective alignment with guidelines, robust governance, continuous performance auditing, and inclusive stakeholder involvement are essential to safeguard fairness. Recommendations address the development of transparent reporting templates for AI outputs, jurisdiction-specific calibration protocols, regular bias audits, and further research into long-term effects on appellate trust and legitimacy.