Implementation Guide

    AB 3030 Disclaimer Implementation: Written, Chat, Audio, and Video

    AB 3030 (Health & Safety Code section 1339.75) sets a different disclaimer placement for each channel: the beginning of each written message, throughout chat and video, and spoken at the start and end of audio. Treat insertion as a deterministic step enforced in the execution path, and record per message the scope decision, the channel rule applied, and the disclaimer presented.

    Where in the message path to insert and verify

    1. 1

      Prompt template

      Easy to implement, but the model can omit, reword, or reposition the disclaimer. It should not be the only control.

    2. 2

      Model output

      Behavior varies across models and versions, so output-level wording is not a consistent control.

    3. 3

      Post-processing

      Deterministic insertion works here, but later edits, templating, or reformatting can displace or strip the text.

    4. 4

      Delivery layer

      Sees the final rendered message, so it is the right point to verify placement. It needs per-channel logic and session-level checks.

    Evidence to retain for each AI-generated message

    • Message identifier, timestamp, channel, and the model and application that produced it
    • The scope decision and the inputs behind it, including the covered entity
    • Review status: whether a licensed provider reviewed the final content, and the version reviewed
    • The channel rule applied and the disclaimer text and human-contact instructions presented
    • Observed placement: beginning of message, persistent display confirmed, or spoken start and end
    • The enforcement action taken (allow, redact, block, escalate) and the policy version in force

    What the statute requires and who it covers

    AB 3030 added Chapter 2.13, Artificial Intelligence in Health Care Services, to the Health and Safety Code, with the operative provision at section 1339.75. It applies to a health facility, clinic, physician's office, or office of a group practice that uses generative AI to generate written or verbal patient communications pertaining to patient clinical information. As of November 2026, the section is codified and the current code text shows the placement rules described below.

    The disclaimer must indicate to the patient that the communication was generated by generative artificial intelligence, and it must come with instructions for reaching a human provider. The statute also provides an exemption for communications that a licensed provider has reviewed. Enforcement runs through the licensing enforcement articles that apply to licensed health facilities and to licensed clinics, so each covered entity should know which licensure applies to it.

    The key design point is that placement is keyed to the interaction type, not to the model or application that produced the text. A rule that holds across every model and every application therefore has to live outside any single prompt.

    Channel map: statutory rule to implementation pattern

    Each row pairs the statutory placement with a concrete pattern. The pattern column is a design interpretation of the text. The statute does not prescribe rendering details such as font, display duration, or exact voice wording, so those are organizational decisions that should be documented.

    • Written: letters, emails, occasional messages: Rule: prominent disclaimer at the beginning of each communication. Pattern: a fixed disclaimer block with human-contact instructions as the first element of the body, checked per message.
    • Chat, including chat-based telehealth: Rule: prominently displayed throughout the interaction. Pattern: a persistent banner or pinned element owned by the session UI, with a session-level check in addition to per-response checks.
    • Audio: Rule: provided verbally at the start and the end. Pattern: a scripted spoken disclaimer before any generated content, and a scripted closing disclaimer triggered by every call termination path.
    • Video: Rule: prominently displayed throughout the interaction. Pattern: a persistent on-screen overlay tied to the video session, with confirmation that it remained displayed.

    Decide when the disclaimer is required

    Before placement comes the question of whether a disclaimer is required at all. Express the decision as explicit conditions that evaluate for every message: the sending entity is a covered entity type, generative AI produced the content, the content pertains to patient clinical information, and no licensed provider has reviewed it. The statute defines patient clinical information, and that definition should be encoded as the scope rule so that administrative messages and clinical messages are separated by a documented test, not by guesswork in a prompt.

    The review exemption deserves the most care. A review that happened before a later edit does not describe the content that was actually delivered. A conservative design ties the review record to the final content, and treats any message without such a record as requiring the disclaimer. That is a design choice, not a statutory instruction, but it avoids the case where an edited message is exempted on the strength of an earlier review.

    Whatever the outcome, record the decision and the inputs behind it. A message that was correctly judged out of scope is only defensible if the judgment is on file.

    Enforce at runtime, not after the fact

    Manual review and periodic sampling find omissions after patients have already received the message. A runtime control instead evaluates each output before delivery and acts on the result: allow a message that carries the correct disclaimer in the correct position, redact or correct content where policy permits, or block a message that lacks it.

    Trussed enforces policy in the execution path across models, agents, and tools, with actions that include allow, redact, block, and escalate for human approval. For AB 3030, that means a policy can require the disclaimer on in-scope AI output and stop output that does not meet it, independent of which model produced the text. Enforcement is at the level of individual interactions, so for chat, video, and audio the persistent display or spoken placement should also be verified in the application that renders them. Trussed runs in production with healthcare organizations, and the evidence generated by enforcement maps to HIPAA.

    None of this settles legal sufficiency. It gives a compliance team a consistent technical control and a record of how it behaved.

    Disclaimer placement by channel

    Written

    Prominent, at the beginning of each communication

    Chat

    Prominently displayed throughout the interaction

    Audio

    Spoken at the start and the end

    Video

    Prominently displayed throughout the interaction

    Review your disclaimer controls with a specialist

    Walk through your channels, scope conditions, and evidence requirements, and see how runtime enforcement and audit-ready evidence work in the execution path.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.