When someone dies by suicide, the focus naturally turns to mental health: depression, isolation, a cry for help that went unanswered. But what if the warning signs were hiding somewhere less obvious, like a courtroom, a police report, or a run-in with the law?

That’s the question a team of USC researchers set out to answer. A new study co-led by the Thomas Lord Department of Computer Science at USC Viterbi and the USC School of Advanced Computing and the USC Suzanne Dworak-Peck School of Social Work is using AI to sift through hundreds of thousands of anonymized death records, looking for a pattern the field has largely ignored: the role of legal interactions in the lead-up to suicide.

“Uncovering Intervention Opportunities for Suicide Prevention with Language Model Assistants”collects and analyzes third-party information about suicide deaths in an effort to reduce them. The research was led by a team including Swabha Swayamdipta, a WiSE Gabilan assistant professor of computer science at the Thomas Lord Department of Computer Science, as well as associate director of the USC Center for AI in Society. Swayamdipta worked alongside Jonathan May, a research associate professor of computer science at the Thomas Lord Department of Computer Science and the director for theArtificial Intelligence Division at the Information Sciences Institute.

The paper was recently accepted to the main conference at the 2026 Association for Computational Linguistics annual meeting this July.

The Viterbi team worked in close collaboration with the USC Suzanne Dworak-Peck School of Social Work. Dworak-Peck’s role in the study was led by John Blosnich, an associate professor and director of the Center for LGBTQ+ Health Equity as well as interim co-director of the USC Center for AI in Society.

The Legal Interaction No One Was Tracking

Specifically, this study focused on legal interactions, which the Viterbi and Dworak-Peck team considered an underrepresented potential cause of eventual suicide. They used the unstructured data found in the NVDRS database to infer and label cases with legal interactions. 

“We don’t typically consider all these life circumstances that the victim goes through in the days before the suicide or in the weeks or months leading up to the suicide,” Swayamdipta said. “A lot of these life disruptors that also lead to suicide have a legal interaction component. We wanted to determine, going through all this unstructured data, how often does that happen?” Eventually, the findings could inform the development of future interventions.

The study is based on  data collected by the National Violent Death Reporting System (NVDRS), a CDC program which compiles info regarding violent deaths in the United States. This program contains a lot unstructured data, and the Viterbi and Dworak-Peck team’s goal was to organize and analyze this data to help understand when suicides may occur. 

The 10% No One Was Looking For

The team was able to find that of the 200,000+ cases in the NVDRS, about 10% of victims had either an implicit or explicit legal interaction before they died by suicide.  This opens up potential opportunities for research into building suicide interventions for legal professionals, especially for clients not getting mental health care.

Researching a sensitive subject like this requires a careful balance of human and AI-driven research analysis. The team settled on a system that featured a first round of human-driven research into a small sample of individual cases to find inferred examples of legal interaction. AI was then implemented based on their research, in order to prevent emotional harm to these researchers and to help sift through this massive pile of data. In other words: this was a three-phase process, going from wholly manual research to human-assisted AI research and finally to unassisted AI research.

“We also did not recommend that anybody on our team look at too many of these cases every single day, and we had checks and balances in place to prevent emotional harms that the study itself might cause to students,” Swayamdipta said. “So we went at a slower pace, but we were manually looking for these indicators of legal interaction in the reports. Once we had that set, then we could use AI to do the same task and then cross check how well it does against our own manual effort.”

This balance between AI and human checks maintained the safety of student researchers while staying focused on its central goal: identifying the link between legal interactions and those who died by suicide. With this efficient and safe system of data analysis, this study could be just the beginning.

“We’re just scratching the surface,” Swayamdipta said. “There is a lot more to do here with the help of AI.”

Published on April 21st, 2026

Last updated on April 21st, 2026