Advancing suicide prevention with human-in-the-loop GenAI

The growing suicide crisis

Suicide remains one of the world’s most urgent public health challenges, accounting for more than 720,000 deaths each year. At the same time, the demand for mental healthcare continues to outpace the available workforce, leaving millions of people without timely access to support.

The latest data from the Substance Abuse and Mental Health Services Administration (SAMHSA) shows that in 2025, 13.9 million U.S. adults (5.3%) experienced serious thoughts of suicide, 4.9 million (1.8%) made a suicide plan, and 2.3 million (0.9%) attempted suicide. An estimated 1.9 million adults experienced all three: suicidal thoughts, a suicide plan, and a suicide attempt.

These trends are moving in the wrong direction. Between 2021 and 2025, the percentage of adults who made a suicide plan increased from 1.4% to 1.8%, while suicide attempts rose from 0.7% to 0.9%. Young adults continue to be disproportionately affected. In 2025, 10.8% of adults ages 18 to 25 reported serious thoughts of suicide, the highest rate of any adult age group.

Despite growing need, access to mental healthcare remains limited. More than 169 million Americans live in federally designated Mental Health Professional Shortage Areas. Among the 54.6 million adults who experienced any mental illness in the past year, only 50.9% received mental health treatment.

These numbers point to an urgent need for new ways to expand access to high quality mental healthcare. Technology alone is not the solution, but thoughtfully designed, clinically supervised AI has the potential to help identify people earlier, extend the reach of clinicians, and connect more individuals with appropriate care when they need it most. In the context of suicide prevention, that has the potential to save lives.

GenAI and suicide risk detection

More people than ever are turning to GenAI for mental health support. In fact, Harvard Business Review found that therapy and companionship were the top use cases for GenAI in both 2025 and 2026.

As GenAI becomes part of how people seek mental health support, safety becomes increasingly important. General purpose AI tools were not designed to deliver clinical mental healthcare, and research has shown they often fail to appropriately detect and respond to suicidal ideation. One study evaluating 29 AI chatbots, including mental health specific applications, found that none met acceptable standards for identifying and managing suicide risk.

Because of this, experts increasingly recommend human-in-the-loop (HITL) models as the gold standard for AI-supported mental healthcare. In a HITL model, AI can identify potential risk, but licensed clinicians review situations that require clinical judgment or crisis intervention and step in when appropriate. This approach helps ensure that people experiencing suicidal thoughts receive timely human support while allowing AI to expand access to care between clinical interactions.

A human-in-the-loop model for suicide risk detection

Pelago’s approach to suicide risk detection builds on years of experience implementing screening and clinical escalation in virtual care. In 2022, Pelago worked with the National Institute of Mental Health’s Ask Suicide-Screening Questions research team to adapt a suicide risk screening pathway for a virtual addiction clinic. In the resulting study, 100% of 252 eligible members completed screening, with 7.5% screening positive for suicide risk. The study demonstrated that universal suicide risk screening could be successfully integrated into virtual care without overburdening clinical workflows.

Sona builds on this clinical foundation for the GenAI era. Pelago designed Sona, an clinical AI specialist, specifically for behavioral healthcare. Unlike general purpose GenAI tools, Sona was built with clinical oversight and a human-in-the-loop (HITL) safety model from the outset. Sona serves as an AI front door to care, continuously assessing acuity, context, and comorbidities to help guide each member to the appropriate level of support while escalating to clinicians when needed.

In a recent safety study under review at JMIR evaluating Sona’s suicide risk detection and escalation protocol, 624 participants completed a pretreatment suicide screening. Of those participants, 81 (13.0%) showed suicidal ideation on the PHQ-9. 31% of those who screened positive were reached by a clinician by phone within an average of 11.2 minutes for additional assessment, while two participants responded to email outreach. Following clinical review, no participants required emergency suicide intervention, and 72% of those who screened positive reported they were already receiving concurrent mental health treatment.

Results from Sona’s safety study

The study demonstrated that proactive, universal suicide risk screening using an empirically validated measure is feasible within an AI-native mental healthcare platform. It also showed that a HITL model can successfully identify individuals experiencing suicidal ideation and connect them with timely clinical follow-up. As GenAI becomes a larger part of how people access mental health support, these findings reinforce the importance of pairing AI with clinician oversight to help ensure safety.