AI transparency and chatbot safety: put controls into the interaction

A disclosure, a machine-readable mark and a safety response are different controls. Identify what the service does, who can use it and which organization must implement each control before relying on a generic AI notice.

Last reviewed

Classify the experience before selecting the notice

Distinguish a support assistant, a companion that sustains a social relationship, a public content generator and a clinical service. Record intended and reasonably foreseeable audiences, whether minors can use it, what the service claims to do and whether outputs influence access to services.

Different laws use different definitions. A chatbot may also need a privacy, employment or medical-device assessment. A disclosure saying “AI” does not by itself resolve those questions. For decisions about individuals, start with the US governance guide or UK automated-decision guide.

Article 50 · Applicable since August 2, 2026

Separate interaction disclosure, content marking and publication duties

The Commission's guidance explains distinct duties for providers of interactive or generative systems and deployers using covered outputs. These include informing people about AI interaction, making synthetic content technically detectable, and disclosures for deepfakes and certain public-interest text. Exceptions depend on the provision and context. Commission transparency guidance.

December 2, 2026 is a limited transition: certain systems already on the market before August 2 have until then to comply with Article 50(2)'s synthetic-content requirements. It is not a general extension of every Article 50 obligation. The new prohibitions on specified sexual deepfake and child sexual abuse material systems also begin December 2. Official implementation timeline.

Recommended practice: decide who implements each notice or marking mechanism, test it on the actual output format, and record what happens when content passes through another system. Keep a rationale for any exception rather than assuming that human editing removes all obligations. The EU guide explains provider and deployer roles.

Enacted safeguards · In force in 2026

California: companion-chatbot protections need their own review

California's SB 243 addresses companion chatbots. The state's 2026 implementation announcement highlights disclosures to minors that the chatbot is not a person and safety protocols addressing self-harm. The signing announcement identifies companion-chatbot protections separately from broader online-safety measures. Governor's 2026 implementation summary; enactment announcement.

Confirm the statutory definition, exceptions, age-related triggers and any separate reporting dates before applying a companion-chatbot checklist to a general business assistant. This overview uses the official implementation summaries; it does not set out a complete SB 243 compliance checklist. SB 243 text and history.

Recommended practice: give a named team responsibility for age-related controls, disclosure placement, escalation and evidence of safety testing. Keep the assessed experience and release version together. An age statement in terms of service is not evidence that a technical safeguard works.

Enacted law · Effective January 1, 2027 · Rules still draft at review

Colorado: a separate law for public conversational AI services

HB26-1263 concerns covered services made available to the public that primarily simulate human conversation. From January 1, 2027, it establishes operator duties including age estimation, AI disclosures, safeguards for known minors, privacy/account tools, self-harm response protocols and reporting. It also restricts claims that outputs are equivalent to specified licensed professional services. Enacted bill summary and signed text.

The Attorney General held the formal rulemaking hearing on October 26. At this review, the official page still describes proposed rules and an October interim draft. Final reporting or implementation details must be checked when published. Colorado rulemaking record.

This law and Colorado's consequential-decision ADMT law address different questions. A service can require both assessments. Read the ADMT guide.

Review professional claims as well as disclosure

Texas requires disclosures for covered governmental AI interactions and AI used in healthcare service or treatment. The healthcare timing includes a specific emergency exception. These are targeted provisions rather than a universal notice duty for all private-sector chatbots. Texas Business & Commerce Code section 552.051.

California's AB 489 also addresses AI falsely presenting itself as a licensed healthcare professional. A bot persona, promotional claim and clinical function should be reviewed together. California implementation summary; medical AI and state requirements.

Keep evidence of the behavior people actually encounter

  • Notice: retain the displayed wording, placement, language and release date.
  • Marking: record the responsible supplier, supported formats and results of checking the mechanism.
  • Safety: document the escalation owner, tested scenarios and corrective decisions.
  • Change: revisit controls when audience, model, persona, distribution channel or intended use changes.

This is a recommended evidence structure. Liquid Learn can connect controls and evidence references to the service review. It does not provide age assurance, content marking, live content moderation or crisis intervention. Those mechanisms must be implemented and validated by the responsible operator and suppliers.

Next editorial check: November 30, 2026, before the December EU milestone and January Colorado commencement. This guide covers selected EU and US provisions; it is not an inventory of every online-safety law.

Connect the review to a real service

See how Liquid Learn connects AI service maps, accountable owners, controls and evidence. Or start with the 24-question health check for an immediate result and downloadable PDF roadmap.

Educational information, not legal advice. Applicability depends on jurisdiction, organizational role and use case. Confirm your obligations with qualified counsel. Liquid Learn supports governance work; using it does not certify or guarantee compliance.