Spring Health's VERA-MH Framework Expands for Mental Health AI Safety Standards

Spring Health's VERA-MH Expansion: Elevating AI Safety in Mental Health

In a groundbreaking move, Spring Health, a global leader in mental health solutions, has extended its VERA-MH framework to include the Harm-From-Others rubric. This framework represents the first open-source AI safety benchmark focused on mental health, now enhanced to evaluate the responses of AI systems when individuals report risks of physical or sexual violence from others.

As the demand for AI-assisted emotional support continues to rise, establishing rigorous standards for evaluating these AI systems has become increasingly essential. The VERA-MH framework was developed through collaboration with mental health professionals, experts on interpersonal violence, and individuals with personal experience in these matters. Together, they crafted a rubric that inspects how AI interacts with users who express concerns about their safety.

The expanded rubric focuses on five critical areas:
1. Risk Detection: Identifying potential dangers reported by users.
2. Confirmation: Verifying claims to ensure the right steps can be taken.
3. Guidance to Human Care: Directing users towards professional help when necessary.
4. Supportive Conversation: Engaging users in a dialogue that validates their feelings and experiences.
5. AI Boundaries: Demonstrating the limitations of AI capabilities in providing help.

Notably, the VERA-MH framework is supported by 100 realistic personas that simulate diverse real-life scenarios, making it a unique and essential tool for assessing AI’s role in mental health care. Dr. Mill Brown, Chief Medical Officer at Spring Health, emphasized the urgent responsibility AI must uphold in ensuring user safety. "In clinical settings, we prioritize established protocols during crises; AI solutions must meet the same standard to protect individuals effectively."

This initiative comes as VERA-MH has received independent validation. The recently published MentalHealthBench paper by OpenAI acknowledges the rubric for suicidal ideation (SI) as one of the few frameworks specifically designed for evaluating safe AI practices in high-risk scenarios.

With the upcoming Domestic Violence Awareness Month, public feedback on the Harm-From-Others rubric is now open for contributions over the next sixty days. This open dialogue will follow a similar process used previously for the suicide-risk benchmark, allowing for community and expert insights to help refine the framework.

The VERA-MH repository is accessible for anyone, including organizations developing general-purpose AI models, enabling them to test their systems against established safety metrics. The framework was initially launched in October 2025 in partnership with members from the AI in Mental Health Safety and Ethics Council, comprised of experts from prestigious institutions like Stanford University, Kaiser Permanente, and Harvard Business School.

Spring Health serves millions worldwide, using an AI-native approach to provide mental health support across various life stages and career paths. In a study from 2025, 92% of their members reported noticeable improvement in mental health issues like depression and anxiety. Recent research published in JAMA Network Open indicated that Spring Health's members also recover more quickly compared to those using other services.

In the realm of mental health care, the intersection with technology presents unique challenges, especially concerning user safety. Spring Health's commitment to continuous improvement, transparency, and robust public engagement exemplifies a proactive approach in harnessing technology for better mental health outcomes. As feedback is gathered on the new Harm-From-Others safety rubric, it portrays a promising step towards creating a safer environment for individuals seeking assistance through AI-enhanced platforms.