When someone is in their darkest moment, trusting AI industry self-regulation is unacceptable. Large language models functioning as de facto therapists for millions of Americans cannot operate in a clinical vacuum, and they cannot be graded solely by the companies that built them.
Calling for guardrails is easy. Drafting statutory mechanics that hold up under clinical reality and software engineering is the hard part. Following our call for national mental health safety guardrails on frontier or generalist AI chatbots, Spring Health conceptualized a model legislative framework for discussion with mental health and AI stakeholders and potential introduction in the U.S. Congress and state legislatures: the Safety Assessments, Frontline Escalation, and Mental-Health Accountability for AI (SAFE-MH AI) Act.
The goal is straightforward: establish a binding national baseline built on independent clinical and technical validation, non-negotiable off-ramps to human care when someone is in crisis, and real regulatory accountability when safeguards fail. An individual with a mental health emergency deserves the same safety guarantees regardless of where they reside.
Enforcing guardrails requires objective, clinically grounded ways to measure safety. At Spring Health, we convened the AI in Mental Health Safety and Ethics Council — a coalition of independent experts across ethics, clinical science, and technology — to build and open-source VERA-MH, a clinically grounded evaluation framework designed to hold AI tools to rigorous clinical standards for mental health safety. By opening these rubrics and personas to public feedback and independent validation rather than keeping them behind a proprietary wall, we ensure that safety standards never depend on a single company, and that any independent auditing requirement we advocate for applies to us on the exact same terms as everyone else.
_405x405.avif)
