State Regulation of Artificial Intelligence In Mental Health: Who Is Being Protected—and Why?

Background and Purpose: AI chatbots are increasingly used in mental and behavioral health contexts, both by individuals seeking support and by practitioners incorporating AI into care-related activities. Yet regulatory frameworks remain limited. State legislatures have emerged as the primary sites for AI governance, though few of the hundreds of AI-related bills introduced nationwide explicitly address mental health chatbot use. This study examines how policymakers constructed regulatory problems related to AI chatbots in mental health during legislative deliberations in Utah, Nevada, and Illinois—the first states to enact such laws—and how those problem constructions shaped the provisions of enacted legislation.

Methods: Audio, audiovisual, and written records of legislative committee hearings and floor proceedings were collected for each state. All oral proceedings were transcribed verbatim. Records were organized by state, bill number, event date, committee, speaker, and speaker role. Guided by an interpretive policy orientation, transcripts were coded line by line to identify recurring themes in how problems were constructed. Enacted statutory texts were then analyzed to examine alignment between legislative problem framing and regulatory design.

Results: Problem construction varied substantially across states and corresponded with distinct regulatory approaches. Utah framed regulatory uncertainty as a barrier to innovation, yielding a consumer protection framework with data safeguards and an affirmative defense mechanism for compliant developers. Nevada emphasized risks to vulnerable populations and the limits of automated clinical judgment, enacting a licensing-based prohibition on AI-delivered counseling while permitting administrative support. Illinois constructed the problem as a boundary violation threatening licensure integrity, extending existing credentialing protections to AI systems and establishing a three-tier classification of permitted and prohibited activities. None of the three states addressed gaps posed by general-purpose language models used for therapeutic support.

Conclusions and Implications: These divergent problem definitions carry practical consequences for social work. Permissive approaches may expand access amid workforce shortages but shift accountability toward reactive enforcement, while licensing-based approaches preserve professional boundaries but may foreclose beneficial applications for underserved populations. Social workers engaged in policy practice should attend to how problem construction in AI governance determines which interests receive protection and which risks remain unaddressed.