Does it matter who writes the news? I study this question by looking at gender differences in reporting on violence against women in Italy. Matching the universe of murders with female victims in 2006–2022 to newspaper articles, I construct indexes for two reporting narratives: victim-centered narratives, which focus on the victim and the violence she experienced and point to support resources, and perpetrator-centered narratives, which focus on motives such as jealousy or a fit of rage. I show that female journalists use more victim-centered and fewer perpetrator-centered narratives, exploiting the conditionally random assignment of reporters to crimes within newsrooms to account for selection in the stories covered. Combining reporter assignment with province-level readership in a staggered difference-in-differences design, I show that greater exposure to female reporters after a murder raises calls to the national helpline for violence against women. An online experiment sheds light on the mechanism: victim-centered narratives shift norms around help-seeking. It also reveals potential engagement costs, as these narratives decrease engagement among men but help retain female readers. The characteristics of those who report the news are thus a relevant supply-side determinant of both news content and reader behavior.

Social divides extend beyond race and gender, but less visible dimensions of social background are rarely recorded in data. We study whether judges treat defendants differently when they are socially closer to them. We measure social proximity as the similarity between judges’ and defendants’ last names, using name embeddings trained on individuals grouped by the street where they live, which yields a continuous measure that varies among individuals with similar observable characteristics. Exploiting the quasi-random assignment of more than 1.2 million criminal cases to judges in Wisconsin, we find that judges are less likely to incarcerate socially closer defendants. Using a model of judicial decisions and a machine-learning prediction of recidivism risk, we find that judges assess the risk of socially closer defendants more accurately but also favor them at the average level of risk. Social leniency is driven by victimless crime, and the effect reverses for offenses involving a victim, who generally come from contexts similar to perpetrators, suggesting judges may also have preferences for protecting socially closer victims.

This paper studies racial in-group disparities in Wisconsin, which has one of the highest Black-to-White incarceration rates among all U.S. states. The analysis is motivated by a model in which a judge may want to incarcerate more due to three factors: (1) taste-based preferences about the defendant’s group identity; (2) higher recidivism risk where the defendant is more likely to commit future crimes; and (3) image motives stemming from the defendant being in the same group as the judge. Further, a judge may have better information on recidivism risk due to two factors: (4) becoming more experienced, and (5) sharing the same group as the defendant. We take these ideas to new data on 1 million cases from Wisconsin criminal courts, 2005-2017. Looking at racial disparities between majority (White) and minority (Black) judges and defendants, we find no evidence for anti-out-group bias (1). Using a recidivism risk score that we construct using machine learning tools to predict reoffense, we find evidence that judges do tend to incarcerate defendants with a higher recidivism risk (2). Consistent with judge experience leading to better information on defendant recidivism risk (4), we find that more experienced judges are more responsive in jailing defendants with a high recidivism risk score. Consistent with image motives (3), we find that when the minority group is responsible for most crimes, minority-group judges are harsher on their in-group. Finally, consistent with judges having better information on recidivism risk for same-group defendants (5), we find that judges are more responsive to the recidivism risk score for defendants from the same group when that group makes up a relatively small share of defendants.

This paper studies the visual representation of crime in U.S. news and whether policies aimed at protecting the identity of arrested individuals affect crime reporting. We assemble 3.46 million crime articles published by 1,848 outlets between 2015 and 2024 and measure their images with a face-level pipeline that detects mugshots, infers race, and links the same individual across articles. Benchmarking against FBI arrest records, we show that Black arrestees are over-represented in mugshot coverage relative to their share of arrests, and the gap is concentrated in high-volume, low-severity offenses such as DUI and drug crimes, where publishing any given booking photograph is discretionary. We then exploit staggered state restrictions on mugshot disclosure in a difference-in-differences design. The bans reduce mugshot publication by about a third and compress the over-representation of Black arrestees where it was largest. Next, we investigate whether crime reporting also adjusts to the removal of mugshots. Using the same design, we show that outlets replace booking photographs with other images, cover fewer arrests and more investigations, and also write about crime differently. Finally, we provide preliminary evidence that reducing mugshot use affects crime perceptions by reducing the salience of crime.

Machine learning based decision-support tools in criminal justice systems are subjects of intense discussions and academic research. There are important open questions about the utility and fairness of such tools. Academic researchers often rely on a few small datasets that are not sufficient to empirically study various real-world aspects of these questions. In this paper, we contribute WCLD, a curated large dataset of 1.5 million criminal cases from circuit courts in the U.S. state of Wisconsin. We used reliable public data from 1970 to 2020 to curate attributes like prior criminal counts and recidivism outcomes. The dataset contains large number of samples from five racial groups, in addition to information like sex and age (at judgment and first offense). Other attributes in this dataset include neighborhood characteristics obtained from census data, detailed types of offense, charge severity, case decisions, sentence lengths, year of filing etc. We also provide pseudo-identifiers for judge, county and zipcode. The dataset will not only enable researchers to more rigorously study algorithmic fairness in the context of criminal justice, but also relate algorithmic challenges with various systemic issues. We also discuss in detail the process of constructing the dataset and provide a datasheet. The WCLD dataset is available at .