AI scribes and client consent: the state law map for 2026

There is no single law that answers "can I use an AI scribe in this session?" There are three layers, and a defensible workflow has to satisfy all of them at once: recording and wiretap law, professional ethics codes, and a growing set of AI-specific state statutes aimed at health care and, increasingly, at mental health in particular.

Layer 1: recording consent#

Ambient documentation captures audio, even if only for seconds before transcription. That brings state eavesdropping and wiretap statutes into play. Most states and federal law follow a one-party consent rule: a participant in a conversation may record it. A minority require all-party consent for some or all recordings — commonly cited examples include California, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Pennsylvania and Washington, with nuances in several others (for example, Connecticut and Oregon distinguish by medium or setting).

In a therapy session, one-party consent technically covers the clinician. But telehealth crosses state lines, groups include several people, and family sessions include minors and guardians. The practical rule most organizations adopt is simple: get explicit consent from everyone whose voice will be captured, every time the setting changes.

Layer 2: professional ethics#

Every major behavioral health profession already requires consent before recording, independent of state law:

• APA Ethical Principles (Standard 4.03): psychologists obtain permission from all persons before recording their voices or images.
• NASW Code of Ethics (1.03(i)): social workers obtain informed consent before audio or video recording clients.
• ACA Code of Ethics (B.6.b): counselors obtain client permission before recording sessions through electronic or other means.

These codes are enforced through licensing boards in many states. An AI scribe used without consent is not just a privacy issue; it can be a licensure issue.

Layer 3: AI-specific state laws#

Since 2024, states have begun regulating AI in health care directly. The laws vary widely in what they cover. The ones most relevant to behavioral health documentation, as we read them:

Two patterns stand out. First, states are drawing a line between AI that supports a clinician and AI that acts as one. Ambient documentation that the clinician reviews and signs sits on the support side everywhere we've looked. Second, disclosure and consent are converging as the baseline expectation, even where a statute doesn't strictly require it.

A consent workflow that satisfies all three layers#

Rather than tracking 50 states, most organizations are better off adopting the strictest reasonable standard everywhere and adding state-specific language where required. A workflow we'd defend:

1. 1Written consent at intake, in plain language and in the client's preferred language, describing what is captured (audio, transcribed and then deleted), what is kept (the transcript, for a set window), who can see it, that it isn't used to train AI models, and that the clinician reviews and signs every note.
2. 2Verbal confirmation at the start of each session, captured in the note as a structured field: who confirmed, when, and in which language.
3. 3Every voice consents — group members, family members, guardians, interpreters. A single member's refusal means that member's contributions aren't captured, or the session isn't recorded.
4. 4Opting out is frictionless and doesn't affect care. The clinician switches to dictation after the session, or writes the note by hand.
5. 5Consent is revocable at any time, and revocation stops capture immediately for future sessions.
6. 6Minors and guardians follow your state's minor-consent rules. Where a minor can consent to their own treatment, they consent to recording too.

How Margent handles consent#

Consent is part of the capture flow, not a separate form. Margent won't start ambient recording until the session has a consent record for every participant — written consent on file, plus a one-tap verbal confirmation that is logged with a timestamp and language. Consent templates are configurable per program and state in your note rules, including Illinois-style written disclosures. A client can decline, and the clinician moves to post-session dictation with one tap.

Live audio streams to transcription and isn't stored by Margent; uploaded audio expires within 24 hours. Transcripts are kept for a retention window your organization sets. Margent doesn't train models on customer data, and no PHI is processed before a BAA is signed. Every note is reviewed and signed by a clinician; nothing is ever auto-submitted to the EHR.