Language about gender roles, sexuality, appearance and LGBTQ+ identity appears more frequently in women’s capital trials than in comparable cases involving men, according to research involving Cornell information scientists and Cornell Law School faculty.
The team analyzed records from 137 U.S. murder trials that resulted in death sentences, comparing cases with similar crime characteristics and defendant demographics. Its study, “Fundamentally Unfair: Analysis of Gendered Discourse in U.S. Capital Trials,” combines computational word analysis with close readings of selected passages.
The researchers describe the approach as a way to help lawyers locate potentially prejudicial language for further examination. A word’s presence alone does not establish stereotyping or show that a trial was unfair, and the study does not measure whether that language changed a verdict or sentence.
Comparing similar cases
The records cover 48 defendants classified as women at their trials and 89 men sentenced to death between 1990 and 2022. The women’s group includes 40 cisgender women and one transgender man on death row as of July 2026, plus seven cisgender women who had died or received reduced sentences.
Sandra Babcock and Rachel Goldberg, clinical professors of law at Cornell Law School, assembled the records. Babcock directs the Cornell Center on the Death Penalty Worldwide and has helped defend hundreds of people facing execution over the past 30 years.
The legal records averaged about 450,000 words each. Researchers divided them into 500-word blocks and used natural language processing to identify patterns in words and phrases.
Comparisons accounted for characteristics such as the victims, manner of murder, motives and additional crimes, including kidnapping or rape. The researchers also examined matched groups based on race, parenthood and the defendant’s relationship to the victim.
They compared how frequently terms appeared in one group’s records with their frequency across the wider collection. Manual interpretation then helped distinguish potentially biased descriptions from language relevant to the facts of a case.
Different themes emerged across groups
Terms about gender roles, sexualization, appearance and LGBTQ+ identity were consistently more common in women’s cases. Language about parenting, intimate relationships and spending also appeared more often in women’s records across much of the analysis.
Drugs and addiction, racial tropes, immigration and violence were more prominent themes in men’s cases. The findings show differences in how defendants are described; they do not establish that every reference within a theme was prejudicial.
The work builds on legal scholarship examining expectations about women as mothers and partners. Prosecutors can use perceived failures to meet those expectations, or a woman’s sexual history, to attack her character even when the material has little bearing on the crime, the researchers said.
The study was inspired in part by the Supreme Court’s January 2025 decision in Andrew v. White. The court vacated a federal appeals judgment and sent Brenda Andrew’s case back for further proceedings after rejecting the lower court’s treatment of her due-process claim over highly prejudicial evidence about her sex life and perceived failings as a wife and mother.
A tool to support legal review
Lead author Camille Harris is an assistant professor of computer science at the University of San Francisco and a former postdoctoral researcher in Allison Koenecke’s lab. Koenecke is an assistant professor of information science at Cornell Tech and the Cornell Ann S. Bowers College of Computing and Information Science.
Cornell information science doctoral students Emma Harvey and Andrea Wang also co-authored the study, along with Babcock, Goldberg and Amanda Potts, an associate professor of public and professional discourse at Cardiff University. The research received National Science Foundation support and was presented at the October 12-14 ACM Conference on AI, Ethics, and Society in Malmo, Sweden.
The men’s records were a convenience sample and may not represent all men’s capital trials. The authors also caution that analyzing blocks from the same case can overstate statistical significance, and that their word lists, developed largely around bias against women, cannot capture every form of stereotyping.
Harvey is extending the work toward a tool that identifies potentially biased language at the sentence level. Koenecke emphasized that computational methods should assist, rather than replace, lawyers’ expertise and the detailed human interpretation needed to assess an individual case.
FingerLakes1.com is the region’s leading all-digital news publication. The company was founded in 1998 and has been keeping residents informed for more than two decades. Have a lead? Send it to news@fingerlakes1.com.
