Stanford convening finds big, unresolved gaps in how mental health AI tools get governed

9:00 AM PT · July 31, 2026

Stanford HAI researchers Caroline Yee, Caroline Meinhardt, Michelle Mello, and Jane Paik Kim wrote up the findings of a convening that brought together policymakers, academics, healthcare providers, AI developers, and patient advocates to work through how mental health and emotional support AI tools should be governed, an area where products have moved into daily use faster than regulatory frameworks have adapted. The piece builds on an earlier Stanford study, led by researcher Andrew Myers, that surfaced a more basic problem underneath the policy debate: when human clinical specialists are asked to judge whether a given chatbot response to someone in emotional distress is “safe,” their assessments frequently diverge from one another, meaning there is no settled, shared definition of safety for evaluators, let alone regulators, to work from. The convening’s participants flagged specific structural gaps, including unclear lines of accountability when a general purpose chatbot, rather than a purpose built clinical tool, ends up used for emotional support, and a lack of consensus on what disclosures or safeguards should be mandatory before such tools reach vulnerable users. The authors argue that governance efforts risk building rules on top of an undefined concept of safety, and call for developing more consistent, validated evaluation standards as a prerequisite to more specific regulation, rather than legislating first and figuring out the clinical evaluation question later.