Fair Machine Learning in Criminal Justice: A Literature Review
Year: 2026
Type: article
Abstract: With the increased prevalence of machine learning (ML) methods in socially consequential decisions in a variety of contexts, including criminal justice, the evaluation of fairness is a critical priority. Additionally, as many aspects of criminal justice have been criticized for fairness concerns and systemic bias, it is increasingly relevant to ensure historical discrimination is not encoded into ML models. The aim of this thesis is to analyze how recent literature has implemented fairness evaluations to ML applications in criminal justice through a literature review. Twenty-six papers are examined through their criminal justice applications and their fairness evaluations. Additionally, the geopolitical contexts of the literature are analyzed through the authors' university affiliations and origins of used data. The reviewed papers often feature United States (US)-based authors and more than half of the papers use data collected from the US criminal justice system. They primarily examine ML applications for risk and recidivism prediction. A variety of fairness metrics are defined and applied, with statistical parity, false positive rate disparity, and equal opportunity appearing most frequently. The reviewed literature defines five sensitive variables used for fairness evaluations: race, gender, age, nationality and employment status. Typically, the sensitive variables are divided into common binary categories: Black and white for race; female and male for gender; under 30 and over 30 years old for age; national and foreigner for nationality; and employed and unemployed for employment status. Intersectionality is rarely considered in the fairness evaluations. Common problems related to creating and evaluating fair ML models for criminal justice applications are raised in the reviewed literature. Insufficient data samples are a common issue, linked to several fairness concerns, such as the treatment of sensitive groups with limited data. Another prevalent concern in relation to fairness is the lack of examination or justification of choices made in creating and evaluating models.
Source: Työväentutkimus Vuosikirja
Author: Lehikoinen, Sanna
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212416417