The new harm-from-others rubric evaluates AI performance across five specific competencies, including risk detection, verification, and the ability to guide users toward human clinical care. Developed alongside experts in interpersonal violence and individuals with lived experience, the framework utilizes 100 realistic personas to pressure-test how systems handle high-stakes disclosures. Dr. Mill Brown, Chief Medical Officer at Spring Health, emphasized that AI requires the same rigorous clinical standards as traditional healthcare, rather than relying on reactive measures during a crisis.
In section Releases
Spring Health Opens AI Safety Rubric on Interpersonal Violence
As reliance on artificial intelligence for emotional support grows, Spring Health has expanded its VERA-MH benchmark to address a critical, often overlooked safety gap: how AI models respond when users disclose risks of physical or sexual violence from others.
This open-source initiative arrives as VERA-MH gains traction in the wider research community, recently cited by OpenAI in its MentalHealthBench study for its focus on safety during suicidal ideation. Organizations building general-purpose models are encouraged to utilize the repository to audit their own systems. Ahead of Domestic Violence Awareness Month, the company is soliciting public feedback on the new rubric via an online portal for the next 60 days, continuing a collaborative development process that previously included contributors from Stanford, Harvard, and Microsoft.
Comments (0)
No comments yet. Be the first!