Does SB 1580 apply if my chatbot only serves triage, not therapy?

Yes. The statute reaches any AI system that presents itself as, or fails to disaffirm being, a licensed mental-health professional in interactions with Tennessee residents. Triage-only positioning does not exempt the system from the disaffirmation obligation. The safest posture is affirmative disclosure at conversation open and machine-visible disaffirmation whenever the model output contains a licensure-adjacent term (therapist, counselor, clinician, therapist assistant, etc.).

What is the evidence standard the AG will apply?

The AG has subpoena authority for conversation records and can compel production of turn-level content. The deployer's evidence must show, per turn, whether a licensure-claim signal was present in the output and, if so, whether a disaffirmation banner was served in the same turn. Application logs that store request and response text without the classification and disaffirmation state fail this standard because they require the deployer to re-classify millions of turns retroactively. Per-decision audit records that capture classification and banner state inline satisfy the standard directly.

Does the statute reach out-of-state deployers?

Yes. Any AI system reachable by a Tennessee resident is in scope. Geo-blocking Tennessee IP ranges is one mitigation, but IP-based geo-blocking falls to any user running a VPN and fails to defend the deployer when a Tennessee resident reaches the system while traveling. Most deployers should assume the statute applies to their platform.

How does SB 1580 interact with HIPAA?

HIPAA and SB 1580 apply concurrently. HIPAA governs the confidentiality and integrity of protected health information (PHI) in the conversation. SB 1580 governs the professional-title claims the AI makes about itself. A single AI response can violate both regimes: PHI mishandling under HIPAA and a licensure claim under SB 1580. The audit architecture that satisfies HIPAA's audit control requirements at 45 CFR 164.312(b) also captures the classification and banner state SB 1580 requires, provided the classifier is wired inline at the response layer.

What other 2026 state chatbot laws should we track?

The Orrick 2026 state chatbot analysis covers the wave. Utah's chatbot disclosure law, California AB 1018, New York A00768A, and Colorado SB 26-189 all reach patient-facing conversational AI in different ways. The compliance architecture that satisfies SB 1580 (response-side classification, per-decision audit records, disaffirmation banners) generalizes across the wave with statute-specific classification taxonomies.

Tennessee's AI therapist-impersonation ban is now in force: the enforcement problem for healthcare chatbot deployers

Tennessee Senate Bill 1580 took effect on July 1, 2026 and prohibits AI systems from presenting themselves as licensed mental-health professionals. Tennessee Attorney General enforcement applies. The statute lands in the middle of a 2026 wave of state chatbot laws (analyzed by Cooley in April) that includes similar restrictions in Utah, California, and New York. Digital-health, employee assistance program (EAP), and payer platforms running patient-facing conversational AI face a compliance record problem that starts operating today.

The problem is not detecting one bad response. It is producing evidence that no bad response occurred across millions of conversation turns.

I want to walk through what SB 1580 requires, why prompt-side controls fail this evidence test, and what response-side policy enforcement plus per-decision audit logs produce that an AG inquiry actually accepts.

Mandate

SB 1580 amends Tennessee's Health Care Consumer Right-to-Shop Act to prohibit AI systems from claiming licensure as a mental-health professional in Tennessee. The statute reaches any AI system that interacts with a Tennessee resident and presents itself as, or fails to disaffirm being, a licensed clinical psychologist, licensed professional counselor, licensed clinical social worker, or licensed marital and family therapist.

What the statute prohibits

Three specific behaviors trigger liability. First, direct impersonation: the AI states or implies it holds a Tennessee mental-health license. Second, licensure-adjacent language: the AI refers to itself as a "therapist," "counselor," "clinician," or equivalent title without a machine-visible disaffirmation. Third, failing to correct a user who assumes the AI is licensed: if the user asks "are you a real therapist," the system must produce a clear "no."

Who the statute reaches

The statute applies to the deployer of the AI system, defined as any entity that makes the AI system available to Tennessee residents. This includes standalone mental-health chatbot vendors, digital-health platforms with embedded conversational AI, EAP platforms whose intake tools use LLMs, and payer platforms whose member-facing chatbots handle behavioral health triage. The out-of-state deployer defense fails: any AI system reachable by a Tennessee resident is in scope.

Penalties and enforcement

Tennessee AG enforcement carries civil penalties per violation and injunctive relief. The AG can subpoena conversation records to substantiate a complaint. The evidence request that follows an AG inquiry is specific: produce every conversation turn where the AI system referred to itself with a mental-health title, along with the disaffirmation language served in the same turn.

Compliance gap

Most patient-facing conversational AI deployments today produce no compliant evidence. The gaps are structural.

Prompt-side controls fail the volume test

The common approach is a system prompt: "You are not a licensed therapist. Never claim to be one." This works as an aspiration and fails as a control. LLMs drift under jailbreak pressure, under long conversations where the system prompt falls out of the effective context window, and under prompts that ask for role-play. The deployer that ships prompt-side controls owns the residual failure rate multiplied by conversation volume. A platform running 10 million turns per month at a 0.01% failure rate has 1,000 potential violations per month before any AG inquiry starts.

Application logs fail the evidence test

The AG's evidence request asks for the disaffirmation language served on turns where the AI referred to itself with a mental-health title. Standard application logs record the request and response as text blobs. They record no policy state, no classification of the response, no evidence that a disaffirmation banner was served. Producing the requested evidence from application logs requires re-parsing every logged response with the same classifier the deployer failed to run inline in the first place.

Vendor-embedded AI is invisible

A material share of behavioral-health chatbot deployments run on vendor SaaS tools that embed model calls under the hood. The deployer's environment never sees the prompt, the response, or the classification. The deployer owns the SB 1580 obligation regardless. The vendor's inability to produce evidence is the deployer's evidence gap.

What surviving an AG inquiry requires

An architecture that satisfies SB 1580 produces, for every conversation turn, a record containing the model output, the classification of that output against a licensure-claim taxonomy, the disaffirmation language served (if any), the identity or session context of the user, and a cryptographic signature. The classification is applied inline before the response reaches the user. Responses that trigger a licensure-claim signal are rewritten or paired with a disaffirmation banner at the response layer, before the response returns to the application.

The architectural pattern is response-side policy enforcement. The request enters the AI system, the model generates the response, the response passes through a policy decision point that classifies the content, and the policy decision point either permits, rewrites, or pairs the response before release. Every step produces evidence. The AG's inquiry asks for turns matching a licensure-claim classification; the audit store returns them.

Beyond Tennessee

The 2026 state chatbot wave applies the same architectural pattern across multiple jurisdictions.

Utah, California, and New York

Utah's chatbot disclosure law requires AI systems in regulated professions to disclose non-human status. California AB 1018 and NY A00768A follow similar frameworks. Each statute produces an evidence request identical in shape: prove disaffirmation was served for specific classes of interaction.

Federal exposure

Tennessee's law does not preempt federal action. HIPAA-covered entities using AI in behavioral health settings still face the HIPAA Security and Privacy Rules on the same conversations. The response-side policy enforcement architecture that satisfies SB 1580 also satisfies HIPAA's audit control requirements at 45 CFR 164.312(b) for the AI decision layer.

Colorado SB 26-189

Colorado's revised AI law (effective January 1, 2027 per the Ropes & Gray analysis) reaches HIPAA-covered clinical AI deployers regardless of the HIPAA exemption question. The same response-classification and audit architecture applies. Building for SB 1580 today builds capacity for Colorado's January 2027 date.

DeepInspect

This is exactly what DeepInspect does. DeepInspect sits inline between your users and the LLM APIs they call. For every request and response, it evaluates identity, data classification, and organizational policy, and makes a permit, rewrite, or deny decision before the traffic reaches the user. For Tennessee SB 1580, the response-side classifier tags any output containing a licensure claim, and the policy engine either strips the claim or pairs the response with the disaffirmation banner.

Every decision produces a per-decision audit record containing the classification, the policy version, the disaffirmation state, the model output, and a cryptographic signature. When the Tennessee AG issues a subpoena, the audit store returns the turns matching a licensure-claim classification with the disaffirmation banner state per turn. The evidence is admissible and independent of the application that generated the conversation.

If you are a healthcare, EAP, or payer platform running patient-facing conversational AI and your July 1 readiness depends on system prompts and application logs, let's talk today.