An act relating to regulating the use of artificial intelligence in the provision of mental health services

The purpose statement now explicitly references 'psychological harm, including death by suicide' as the harm the bill seeks to prevent, and simplifies the three-point structure into a single sentence stating services must be delivered by mental health professionals rather than independently by AI systems.

The official version's entirely new 26 V.S.A. Chapter 120 is removed from the unofficial version. This chapter had defined permitted uses (scheduling, billing, transcription with consent, deidentified data analysis), prohibited uses (AI making therapeutic decisions, independently diagnosing, generating treatment plans), confidentiality and consent requirements, and exceptions for religious counseling, peer support, and educational resources.

The core prohibition in 18 V.S.A. section 7115 changes from banning entities that 'represent artificial intelligence as providing therapeutic judgment, diagnosis, treatment, or therapeutic communication' (a misrepresentation standard) to requiring that mental health services be 'provided by a mental health professional' with an exception for approved institutional review board or privacy board studies under 45 C.F.R. section 164.512(i)(1)(i)(A) and (B).

The unofficial version adds a new subsection (d) to 18 V.S.A. section 7115 explicitly permitting mental health professionals to use HIPAA-compliant AI tools, including FDA-authorized or cleared digital therapeutics and software as a medical device products, provided the professional reviews and approves the services and the product is prescribed or recommended by the professional.

The definition of 'artificial intelligence' changes from a functional description ('a machine-based system or software that...is capable of perceiving an environment through data acquisition') to a more standard formulation ('an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs'). The separate definition of 'generative artificial intelligence' is removed.

The definition of 'mental health services' in 18 V.S.A. section 7115 is narrowed from 'support, counseling, therapy, or psychotherapy services provided by a mental health professional to diagnose or treat an individual's mental or behavioral health or provide ongoing recovery support, excluding religious counseling' to 'services provided to diagnose, treat, or address an individual's mental health or behavioral health through therapeutic communications and therapeutic decisions.' The definition of 'therapeutic communication' is expanded to include nonverbal interactions and lists five specific examples (direct interactions, clinical guidance, clinical support including reassurance/empathy, collaborating on treatment plans, and delivering feedback for growth).

The unprofessional conduct cross-reference in 3 V.S.A. section 129a changes from citing 'misuse of artificial intelligence pursuant to 26 V.S.A. section 7101' (the removed Chapter 120) to 'engaging in the prohibited use of artificial intelligence pursuant to 18 V.S.A. section 7115.' Additionally, the unofficial version amends 26 V.S.A. section 1354 (a different professional conduct statute) to add AI misuse as unprofessional conduct, rather than creating the new Chapter 120.

The definition of 'artificial intelligence' in both 18 V.S.A. 7115 and 26 V.S.A. chapter 120 was rewritten from 'an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs' to 'a machine-based system or software that, depending on human objectives, is capable of perceiving an environment through data acquisition and then processing and interpreting the derived information to take an action or to imitate intelligent behavior.' The new definition adds language about learning and adapting behavior.

The definition of 'generative artificial intelligence' was changed from 'an artificial intelligence that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the system's training data' to 'foundation models used in artificial intelligence systems specifically intended to generate, with varying levels of autonomy, content such as complex text, images, audio, or video.' This narrows the focus to foundation models specifically.

The definition of 'therapeutic decision' (subsection (a)(10) of 26 V.S.A. 7101) was removed entirely. This defined it as 'the final clinical determination regarding diagnosis or selection, modification, or termination of treatment or care' and excluded algorithmic risk scoring and clinical decision support tools used under supervision. Its removal is consistent with the shift from 'therapeutic decisions' language to 'clinical responsibility' language.

In 26 V.S.A. 7101(d)(1), prohibited uses were restructured. The word 'independently' was removed from all prohibitions, making them absolute rather than limited to autonomous AI actions. New prohibited activities were added: 'directly interact with clients' and 'detect emotions or mental states.' A new subsection (B) was added prohibiting professionals from offering, providing, or advertising services that represent AI as providing therapeutic judgment, diagnosis, treatment, or therapeutic communication - duplicating the prohibition already in 18 V.S.A. 7115(b).

In 26 V.S.A. 7101(c)(2), the consent requirement was expanded from covering only when AI is used to 'record' identifiable therapeutic communications to covering when AI is used to 'record, transcribe, or process' such communications. A new exception was added: consent is not required for administrative or deidentified operational uses.

In 26 V.S.A. 7101(b), the standard for permitted uses changed from the professional retaining 'sole responsibility for therapeutic decisions' to retaining 'clinical responsibility.' Additionally, 'transcription and documentation support' now requires 'patient or client consent,' and the prior requirement that the professional 'reviews, modifies where necessary, and approves the final product' was removed.

Section 1 (Purpose) now says therapeutic judgment remains the responsibility of 'qualified mental health professionals' rather than just 'mental health professionals,' adding a qualifier that could affect who is considered covered by the act's purpose statement.

The entire new Chapter 120 (26 V.S.A. ch. 120) on 'Artificial Intelligence in Regulated Professions' was deleted. This removed detailed definitions (administrative support, clinical responsibility, consent, supplementary support), permitted-use lists, confidentiality and consent requirements, and exceptions for religious counseling, peer support, and educational resources.

The core prohibition in 18 V.S.A. 7115(b) was broadened from banning services that 'represent artificial intelligence as providing therapeutic judgment' to a blanket ban: no corporation or entity may provide, advertise, or offer mental health services to the public unless they are (1) provided by a mental health professional or (2) part of an approved IRB/privacy board study.

The purpose statement was narrowed from three goals (professional responsibility, individual choice including peer/faith-based options, and responsible AI use for administrative functions) to a single goal: safeguarding individuals from 'psychological harm, including death by suicide' by ensuring services are delivered by mental health professionals rather than independently by AI systems.

A new exception was added in 18 V.S.A. 7115(d): mental health professionals may use HIPAA-compliant AI tools, including FDA-authorized software-based medical products and digital therapeutics, provided the professional reviews and approves the services.

The definition of 'therapeutic communication' was expanded to include nonverbal interactions and now lists five specific examples (direct client interactions, clinical guidance, emotional support including reassurance/empathy, treatment plan collaboration, and feedback for growth). A new definition of 'therapeutic decision' was added. The definition of 'mental health services' was simplified and no longer requires the service be provided by a mental health professional.

Enforcement was shifted from the professional regulation framework (26 V.S.A. 7101) to consumer protection law (9 V.S.A. ch. 63). The unprofessional conduct provision in 3 V.S.A. 129a now references 18 V.S.A. 7115 instead of 26 V.S.A. 7101, and a new provision was added to the Board of Medical Practice's unprofessional conduct section (26 V.S.A. 1354).

A new Section 6 requires the AI Advisory Council to submit a written report to four legislative committees by January 15, 2027, regarding regulation of AI use by mental health professionals, including recommendations for further legislative action.

Four new members were added to the AI Advisory Council in 3 V.S.A. 5023: the Director of Professional Regulation, the Executive Director of the Vermont Board of Medical Practice, the Executive Director of Racial Equity, and the Attorney General.

The prohibition on AI use in Section 7101(d)(1) was rewritten. The old version banned professionals from using AI to make therapeutic decisions, directly interact with clients, generate treatment plans, or detect emotions or mental states. The new version only bans AI that 'independently' makes therapeutic decisions, independently diagnoses, independently determines treatment, or independently generates treatment plans. The explicit bans on direct client interaction and emotion detection were removed.

The consent requirement in Section 7101(c)(2) was narrowed. Previously, patient consent was required when AI was used to record, transcribe, or process identifiable therapeutic communications. Now consent is only required when AI is used to record identifiable therapeutic communications, removing the consent requirement for transcription and processing.

The definition of 'artificial intelligence' was changed from a description focused on perceiving environments and imitating intelligent behavior to one aligned with NIST's framework: 'an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.'

The definition of 'generative artificial intelligence' was changed from referring specifically to 'foundation models' to a broader description: 'an artificial intelligence that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the system's training data.'

A new definition of 'therapeutic decision' was added in Section 7101(a)(10), defining it as 'the final clinical determination regarding diagnosis or selection, modification, or termination of treatment or care' and explicitly excluding 'algorithmic risk scoring, data analytics, or other clinical decision support tools when used under the supervision and authority of a licensed mental health professional.'

The condition for permitted AI use in Section 7101(b) was changed from requiring the professional to retain 'clinical responsibility' to retaining 'sole responsibility for therapeutic decisions,' narrowing the scope of what the professional must be responsible for.