AI Regulation Tracker  /  Regulation enacted

Maine LD 2082: AI Cannot Deliver Therapy, and a Signed Consent Form Now Gates Your Ambient Scribe

Maine just drew a hard line through the therapy room. A licensed clinician cannot route therapeutic communication or treatment decisions through AI, and even an ambient scribe needs the patient's written, revocable consent first. Here is what is binding, what is still allowed, and the intake step you should build now.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published June 27, 2026  ·  Last updated June 27, 2026

Maine put the therapy back in human hands

The tools arrived in the therapy room faster than anyone wrote rules for them. An ambient scribe listens to a session and drafts the note. A summarizer turns an hour of conversation into three bullet points. A triage assistant reads an intake form and suggests where a new client should be routed. None of that looked like a policy question until a clinician asked the obvious follow-up: how far into the actual care is the model allowed to reach.

Maine answered on April 13, 2026, when the Governor approved LD 2082, "An Act to Regulate the Use of Artificial Intelligence in Providing Certain Mental Health Services," now Public Law Chapter 687. The law takes effect July 29, 2026. It does not ban AI from a clinical practice. It bans AI from being the clinician.

The line the statute draws

The core of LD 2082 is a boundary between two kinds of work. The statute first bars AI from being the therapist at all. In its words, "A person may not provide, advertise or otherwise offer therapy or psychotherapy services, including through the use of Internet-based artificial intelligence, to the public unless the therapy or psychotherapy services are provided by a licensed professional." A violation of that prohibition is treated as a violation of the Maine Unfair Trade Practices Act.

The law then sets a hard limit on what AI may do even when a licensee uses it for permitted support. The statute provides that a licensee "may not allow artificial intelligence to: A. Make independent therapeutic decisions; B. Directly interact with clients in any form of therapeutic communication; or C. Generate therapeutic recommendations or treatment plans without review and approval by the licensee." It defines "therapeutic communication" as "any verbal, nonverbal or written interaction, conducted in a clinical or professional setting, that is intended to diagnose, treat or address an individual's mental, emotional or behavioral health concerns," and lists examples that include providing guidance and interventions, offering emotional support, collaborating on therapeutic goals or treatment plans, and offering behavioral feedback. That is the substance of therapy, and the model may not deliver it, may not interact with a client through it, and may not produce a treatment plan or recommendation that a clinician has not reviewed and approved.

On the other side is everything that supports the work without being the work. The statute lets a licensee provide "administrative support or supplementary support in the delivery of therapy or psychotherapy services through the use of artificial intelligence," but only if the licensee "maintains full responsibility for all interactions, outputs and data use associated with the use of artificial intelligence." Administrative support, as the law defines it, covers tasks that do not involve therapeutic communication, such as scheduling, billing, and logistical messages. Supplementary support covers other non-therapeutic tasks like preparing client records and organizing referrals, subject to clinician review. The point is unmistakable. Software can sit beside the clinician. It cannot sit in the chair.

Consent is now a precondition, not a courtesy

Here is the part that reshapes a Monday morning. LD 2082 makes the client's consent a condition of any supplementary AI use, and it ties that use to recording. The statute permits a licensee to use AI for supplementary support "only when the client's therapeutic session is recorded or transcribed, and only if" the client "is informed in writing" of three things, that AI will be used, the specific purpose of the tool, and how session data will be stored, retained, used for training, and deleted, and the client "provides consent to the use of artificial intelligence." Read that against how ambient scribes actually work and the sequence becomes clear. The recording sits inside the same permission as the consent, so the practical posture is to secure the written disclosure and consent as a precondition rather than after the fact.

The consent is not a casual yes. The statute defines "Consent" as a "clear, explicit and affirmative act" that communicates the individual's "written agreement" and that "is revocable by the individual," and it expressly excludes consent obtained through a broad terms-of-use acceptance or through deceptive design. So the consent becomes a real artifact in the chart, written, attributable, and reversible, not a verbal aside buried in a first session. A client who agreed in March can withdraw in June. The law reinforces that the client cannot be punished for saying no: a licensee "may not deny, or refuse to provide" services solely because the client withheld consent to supplementary AI use, and any client waiver of these protections is "void and unenforceable."

What this means for the tools already in your practice

If you have adopted an ambient note-taker, the technology does not have to leave. The intake process has to change around it. Build a distinct AI consent step that a client reviews and signs, separate from the general consent to treat, so the agreement to be recorded and summarized by software is its own decision a client can say no to. Keep a clinician's name attached to every AI output as the reviewer of record, because the statute puts responsibility on the licensed professional regardless of what the model wrote.

Then audit for the one thing the law will not forgive. Make sure nothing in your stack is quietly crossing into therapeutic communication. A chatbot that answers a client between sessions, an assistant that drafts a message of emotional support sent under the clinician's name, a model that proposes a treatment plan a clinician merely signs without independent judgment, those are the uses LD 2082 was written to stop. The safe posture is simple to state and harder to verify: the human does the therapy, the software does the paperwork, and the client agreed in writing before any software touched the session.

Why pending still means act now

The law is not yet in force, and that is exactly why the work belongs on the calendar this quarter. July 29, 2026 is a fixed date, and the change it requires is not a setting toggle. It is a new consent artifact, a revised intake script, a documented review step, and an honest inventory of what your AI tools actually do. Practices that wait until the effective date to find out their ambient scribe has been running without a compliant consent will be retrofitting consent onto encounters that already happened. Building the consent step before the deadline is the difference between a clean start and a cleanup.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.