Mental Health AI Gets Its First Safety Standard for Detecting Domestic Violence Risk

As artificial intelligence chatbots increasingly serve as first-line emotional support for millions of people, a critical question has emerged: how do these systems respond when someone discloses they're at risk of violence? Spring Health is now opening that question to public scrutiny with a new safety rubric designed to evaluate AI performance in domestic violence scenarios.

The mental health company has expanded VERA-MH, the first open-source AI safety benchmark for mental health, to include a harm-from-others safety rubric. The framework evaluates how AI systems respond when adults describe being at risk of physical or sexual violence. Spring Health is inviting 60 days of public comment on the rubric starting September 25, 2026, ahead of Domestic Violence Awareness Month.

Testing AI Against Real-World Violence Scenarios

The harm-from-others rubric evaluates AI responses across five critical areas: detecting risk, confirming it, guiding users to human care, maintaining supportive conversation, and following appropriate AI boundaries. The framework uses 100 realistic personas to test how AI systems handle interpersonal violence disclosures.

This expansion follows the same collaborative process Spring Health used for its suicide-risk benchmark, which launched in October 2025. That initial VERA-MH framework has gained independent validation—OpenAI's MentalHealthBench paper recognized VERA-MH's suicidal ideation rubric as one of the few frameworks built specifically for safe AI risk assessment.

The development comes as 170 million people worldwide now have access to Spring Health through employers and health plans, underscoring the scale at which AI-powered mental health tools operate.

Clinical Expertise Shapes Safety Standards

Spring Health developed VERA-MH in collaboration with its AI in Mental Health Safety and Ethics Council, which includes clinicians and researchers from Stanford University School of Medicine, Kaiser Permanente, Yale School of Medicine, Harvard Business School, and Charité – Berlin University Hospital. Council members also represent major healthcare organizations including JPMorgan Chase, UnitedHealthcare, and Microsoft.

The public comment period allows frontline clinicians with domestic violence experience to shape safety standards before widespread adoption. For health tech companies building AI-powered mental health tools, VERA-MH provides an open-source testing standard that could reduce liability risk and improve clinical safety protocols.

Healthcare providers and health systems can use these benchmarks to evaluate third-party AI tools before implementation—a particularly important capability as AI chatbots become more prevalent in clinical workflows.

Proven Outcomes Drive Safety Innovation

Spring Health's focus on safety standards comes alongside documented clinical outcomes. According to a 2025 study published in the Online Journal of Public Health Informatics, 92% of Spring Health members improved or recovered from depression or anxiety. Research published in JAMA Network Open found members recover 5.9 weeks faster than those using a leading competitor. A Validation Institute study reported 21% net savings in mental health costs and 25% lower physical health spending over two years.

As AI continues to expand its role in mental health care, open-source safety frameworks like VERA-MH may establish the clinical standards that shape both industry practice and future regulatory frameworks for high-risk scenarios.

Sources

1. Spring Health Opens VERA-MH Harm-From-Others Safety Rubric for Public Comment – Health Technology Net