Spring Health releases AI safety rubric for mental health services

Spring Health, a global mental health company, has expanded its open-source AI safety benchmark, VERA-MH, to include a new clinical safety area: "harm from others." This expansion addresses how AI systems respond when an adult describes being at risk of physical or sexual violence from another person. The new rubric has been opened for 60 days of public comment ahead of Domestic Violence Awareness Month.

The VERA-MH Harm-From-Others rubric evaluates an AI's ability to detect risk, offer supportive conversation, and guide individuals toward human care. Developed with clinicians, experts in interpersonal violence, and individuals with lived experience, it uses 100 realistic personas. This marks the first open-source benchmark of its kind, allowing organizations to test the safety of their mental health AI applications.

Spring Health stated that as more individuals turn to AI for emotional support, establishing standards for measuring the safety of these technologies is crucial. Chief Medical Officer Dr. Mill Brown emphasized the need for AI to undergo the same rigor and training as human clinicians. The benchmark's previous rubric for suicidal ideation has already been cited by OpenAI.

The open-source nature of the VERA-MH rubric and its associated personas allows any organization to use and contribute to its development. Spring Health aims to gather broad feedback to further refine the evaluation method for AI in mental healthcare.