In 2001, Brenda Andrew was convicted of murdering her husband, and in 2004 she was sentenced to death.
But 21 years later, the Supreme Court found that prosecutors’ evidence about her sex life, appearance and perceived failings as a wife and mother could be unduly prejudicial, and remanded the case to a lower court.
Inspired in part by this decision, Cornell information science researchers, together with Cornell Law School faculty, found that prosecutors invoke a range of gender stereotypes against women defendants to increase the likelihood of convictions and death sentences. Their research was supported by a grant from the National Science Foundation.
In particular, lawyers use terms associated with gender roles, sexualization, appearance, parenting and LGBTQ+ identity more frequently in the trials of women than of men. The findings could help spur systemic changes in how female defendants are viewed, and talked about, in court.
“Creating methods that can be applied to real cases to identify individual instances of bias is important in correcting injustices in the legal system,” said Emma Harvey, a doctoral student in information science and co-author of “Fundamentally Unfair: Analysis of Gendered Discourse in U.S. Capital Trials,” which was presented at the ACM Conference on AI, Ethics, and Society, held Oct. 12-14 in Malmo, Sweden.
“Our long-term goal, especially in collaboration with the folks at Cornell Law, is thinking about how to address all of these different pernicious forms of biases and taking steps to mitigate them concretely,” said Allison Koenecke, assistant professor of information science at Cornell Tech and the Cornell Ann S. Bowers College of Computing and Information Science.
The lead author is Camille Harris, assistant professor of computer science at the University of San Francisco and a former postdoctoral researcher in Koenecke’s lab. Harvey and Andrea Wen-Yi Wang, also a doctoral student in information science, are co-authors.
The team employed natural language processing (NLP) to analyze thousands of pages of transcripts from legal proceedings involving defendants in murder trials.
Co-authors Sandra Babcock and Rachel Goldberg, clinical professors of law at Cornell Law School, compiled the corpus of U.S. capital trial transcripts – 137 in all, including those of all cisgender women currently on death row. Babcock is faculty director of the Cornell Center on the Death Penalty Worldwide, and over the last 30 years has helped defend hundreds of people facing execution around the world.
“Our research findings draw on the power of NLP and the nuance captured by qualitative coding,” Babcock said. “Our project illustrates how scholars in disparate fields can work together to produce rigorous, multi-layered analysis of voluminous legal texts.”
Babcock is among a handful of legal scholars who’ve explored the idea of gender bias in capital cases. One common theme: the false concept that women possess an intrinsic maternal instinct leads to harsh judgments of their parenting skills. Similarly, perceived promiscuity is often used against a female defendant, while men’s consensual intimate relationships are almost never scrutinized.
For this work, Koenecke’s team analyzed the transcripts of 48 women and 89 men sentenced to death between 1990 and 2022. The transcripts averaged around 450,000 words, so the researchers “chunked” the text into blocks of 500 words each to facilitate NLP analysis and finding key words and phrases.
The researchers matched trials of women and men based on similar factors in the case: the nature of the victim or victims, manner of the murder, motives and additional crimes committed such as kidnapping or rape. The researchers evaluated patterns of gender-associated key terms within the texts, to understand how language differed when used across women’s and men’s trials.
They then calculated the frequency of a given word in a document, compared to how frequently it’s used in the entire corpus of cases, to determine whether certain words or phrases are used more frequently in trials of one gender versus the other.
The researchers found that, indeed, themes of gender roles, sexualization, parenting, intimate relationships and spending consistently appeared more frequently in women’s transcripts than in men’s. Themes found more commonly in men’s trials included drugs/addiction, racial tropes, immigration and violence.
Koenecke said what surprised her most about the findings is that such a simple NLP method could detect these biases.
“When these sorts of naive analyses are combined with qualitative analyses, you can actually go a long way in understanding real stories about where this sort of gendered discourse might appear in capital trials,” she said. “I think this speaks to the power of even the most basic of computational tools.”
And combining these basic NLP methods with more qualitative analysis – including reading and interpreting the transcripts manually – “makes even simple methods much more powerful,” Harvey said.
Harvey is extending this research, working on a tool that can identify, at the level of individual sentences within a document, evidence of biased verbiage toward a defendant. But computational methods can’t replace attorneys’ expertise in analyzing and identifying gender-based bias, Koenecke said.
“What we want to do is use this range of computational methods to help identify points of biases, without replacing the real human expertise that’s needed,” she said. “All of our work really centers the expertise of the actual human lawyers, who’ve done the qualitative annotating that’s necessary to better understand what’s actually happening in the courtroom.”
Amanda Potts, associate professor of public and professional discourse at Cardiff University, also contributed to this research, which was supported by a grant from the National Science Foundation.
