Picture a forlorn man walking into a lawyer’s office after his marriage has collapsed. The bank is threatening foreclosure, and his finances are in tatters.

His problems may be legal on paper, but in person they can look far more complicated: panic, shame or the numb resignation of someone who can no longer imagine a way out. He did not come for counseling, and his lawyer is not a therapist. Still, the attorney may be sitting beside him at the worst moment of his life — with a chance to help before despair slides into tragedy.

That overlooked encounter sits at the heart of research by Swabha Swayamdipta, WiSE Gabilan Assistant Professor of Computer Science in the Thomas Lord Department of Computer Science at  USC Viterbi and the Stevens School of Computing and AI, her computer science Ph.D. student and paper lead author, Jaspreet Ranjit, and John R. Blosnich, as associate professor at the USC Suzanne Dworak-Peck School of Social Work and interim co-director of the USC Center for AI in Society (CAIS).

“What we have shown is that legal professionals interact with people who later die by suicide much more often than we previously knew,” Ranjit said. “You can’t design or evaluate any type of intervention until you know where the opportunities exist.”

Their paper, “Uncovering Intervention Opportunities for Suicide Prevention with Language Model Assistants,” was presented at the Association for Computational Linguistics’ 64th annual meeting in San Diego in July 2026. Co-authors include Jonathan May, research associate professor of computer science and principal scientist and AI division director at  USC’s Information Sciences Institute;  Hyundong J. Cho, May’s former Ph.D. student and now a member of the technical staff at Intology; Claire J. Smerdon, a software engineer at Palo Alto Networks who holds a USC Viterbi master’s degree in computer science; USC Viterbi computer science alumnus Yoonsoo Nam; and Myles Phung. 

Finding what the reports did not label

The researchers studied the Centers for Disease Control and Prevention’s National Violent Death Reporting System. Its records combine standardized categories with written accounts drawn from coroners, medical examiners and law enforcement. The narratives can describe the circumstances before a death in detail, but the database has no standard category for contact with a lawyer.

Most of that contact was hidden in plain sight. A report might mention divorce proceedings, a custody fight, bankruptcy, eviction or an arrest without naming an attorney. A reader can infer that a legal professional was probably involved. A simple word search cannot.

To find those hidden connections, Ranjit and her colleagues built an AI assistant that could interpret context instead of simply searching for words such as “lawyer” or “attorney.” Public health experts first taught the system what qualified as an explicit, implied or absent legal interaction. The AI then reviewed sample narratives and assigned each one a label. Whenever it made a mistake — missing a likely legal interaction or finding one that was not there — the experts corrected it and refined the instructions. They repeated that process until the system could reliably analyze hundreds of thousands of records, with human experts continuing to set the rules and check its work.

An earlier version of the study examined 270,000 cases and found evidence of legal contact in 10.4%, or about 28,000. The final paper expanded the data to 756,000 cases and found evidence in 12.7%, about 96,000 narratives. In other words, evidence of legal contact appeared in about one of every eight cases, even though only one in 250 — 0.4% — explicitly mentioned a lawyer.

From discovery to a careful intervention

The study does not prove that intervention by lawyers will prevent suicide, and it does not ask attorneys to become counselors. It identifies a place where people in crisis already appear and where a trained professional might notice what others have missed.

“This gives us the tools to try to see if there is a reasonable likelihood that intervention, or that screening, is useful enough,” May said.

Added Swayamdipta: “This is a nice reminder of  what happens when humans interact with certain kinds of technologies and how these technologies can be used for changing human lives. This is going to make a real-world impact.”

One possibility is to train lawyers to recognize signs of suicide risk, respond supportively, and connect distressed clients with appropriate help. Ranjit suggests that law schools incorporate Mental Health First Aid (MHFA) training into their curricula, which teaches non-specialists how to recognize and respond when someone is experiencing a mental health problem or crisis. When appropriate, lawyers could also administer a short intake questionnaire to clients facing divorce, bankruptcy, eviction, custody disputes, or criminal cases. Any such screening tool would need privacy safeguards and evidence that it actually helps.

Lawyers may be only the first newly visible link in a larger potential safety net. If the researchers are right, the next chance to save a life may be waiting in a bank, shelter or courthouse — where someone in despair has gone for help with a different kind of emergency.

“Our next step is: Can we do this with the financial sector?” Ranjit said. “What does that look like with loan default or mortgage foreclosure? Can we use additional data sources to find evidence for new interventions?”

Published on September 29th, 2026

Last updated on September 29th, 2026