Stanford HAI mapped regulatory gaps in mental health AI

The design

On 24 July 2026, Stanford's Institute for Human-Centered AI (HAI) published 'The Complexities of Governing Mental Health AI,' an analysis by Caroline Yee, Caroline Meinhardt, Michelle Mello and Jane Paik Kim summarizing a convening of 'policymakers, academics, healthcare providers, AI developers, and patient advocates' held under the Chatham House Rule. The institute's own published piece is authored by its policy research manager and faculty affiliates in law, health policy and psychiatry, and is framed as the group's own analysis rather than a Stanford institutional position or enacted policy.

What the evidence says

The piece names three gaps its convening identified: first, that 'there is no broad consensus on what specifically counts as an AI mental health tool,' illustrated with named examples including 'Character.ai's Trauma Therapist' and general-purpose chatbots that 'assume a particular persona or role, such as that of a therapist, friend, or trusted partner'; second, that evaluation methods for chatbot safety are lagging behind what multi-turn conversations require; and third, that near-term policy fixes should proceed alongside deeper structural questions. A separate HAI news article the same period reported a Stanford-affiliated empirical study, distinct from this policy piece, that surveyed 1,131 Character.AI users recruited through Prolific and found intense companion-motivated chatbot use correlated with lower well-being among users with smaller offline social networks, a correlational finding the researchers themselves stated does not establish that chatbot use causes reduced well-being.

What it asks of people

The governance piece asks policymakers and practitioners to first agree on definitions before writing rules, pointing out that a clinically developed therapy chatbot and a general-purpose assistant a user turns to during a crisis currently sit in the same regulatory blind spot. It asks AI developers to close the gap between how quickly products ship and how slowly evaluation methods for multi-turn emotional conversations have developed, a critique the authors direct at the industry's current practice rather than at any single named company.

Privacy and safeguards

The piece does not itself propose a specific data-privacy rule; its scope is the definitional and evaluation gaps in mental-health AI governance, and it frames these explicitly as unresolved 'for further research' rather than settled recommendations. It draws a boundary worth preserving: tools that 'provide mental health support ... but stop short of making medical claims' currently fall, in the authors' own words, 'outside the regulatory scope of medical devices,' a gap the authors identify rather than endorse.

• Which of the three named gaps has since been addressed by a specific state or federal rule, and which remain open?
• Does the Character.AI user study's correlational finding get restated elsewhere as a causal claim about chatbot use and loneliness?
• How does the Chatham House framing of this convening affect what can be attributed to which named author?

Stanford HAI's own account is a snapshot of an unresolved policy debate among named participants, not a completed regulatory framework; treating its three identified gaps as open questions, rather than as HAI's settled recommendations, matches how the piece itself is framed.