Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
AI Chatbot Disclosure: What California, Colorado and the EU Require
The short answer
Every regime CompanionGuard covers asks for the same core act — tell the user they are talking with an AI — but they differ in when and how. California's pack: disclose at first interaction and on request. Colorado's pack: disclose AI nature clearly. The EU AI Act's Article 50 transparency duties have applied since 2 August 2026. CompanionGuard's free disclosure snippet carries the banner for one bot; the runtime tracks disclosure state per session on Pro and Team. This is a toolkit's summary, not legal advice.
What is AI disclosure for a chatbot?
AI disclosure is the act of telling a user, clearly, that they are interacting with an AI system rather than a human. In CompanionGuard it is implemented as an embeddable disclosure snippet with jurisdiction banner wording, a statement on the hosted safety-protocol page, and — in the runtime — a per-session disclosure state that must be satisfied before ordinary conversation proceeds.
Create the free safety kit Next guide All guides
CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.
Three regimes, one core act, different triggers
| Regime | Disclosure duty, as the product states it | Timing |
|---|---|---|
| California — SB 243 | Disclose AI nature at first interaction and on request | Operative 1 Jan 2026 |
| Colorado — Chatbot Safety Act | Disclose AI nature clearly to users | Joins 1 Jan 2027 (per secondary sources — verify enrolled text) |
| EU — AI Act Article 50 | Transparency duties for AI interaction | Applied since 2 Aug 2026; legacy grace period ends 2 Dec 2026 |
California's formulation is the most operationally specific: two trigger points, the start of the interaction and any later request. A banner shown once at signup arguably meets 'clearly' but fails 'on request' if the bot, asked directly 'are you a real person?', answers in character. Colorado's 'clearly' is a quality standard rather than a trigger list. The EU duty, as the product's llms summary puts it, is a transparency duty that has applied since August 2026, with the grace period for legacy systems ending 2 December 2026 — the nearest deadline of the three for builders who have not yet moved.
The banner wording the packs ship
Both US packs use the same banner skeleton: 'You are talking with {bot}, an AI companion — not a human. If you are in crisis, call or text 988 (Suicide and Crisis Lifeline) any time, free and confidential.' Three properties of that wording are deliberate. It names the bot, so the disclosure attaches to this conversation, not to a brand in the abstract. It denies humanity in plain words — 'not a human' — rather than hiding behind 'virtual assistant' ambiguity. And it fuses the crisis line into the disclosure itself, so the two most important sentences a vulnerable user can read arrive together, at the start.
The minor variants differ by state: California's pack adds 'we will remind you every 3 hours, as California law requires for known minors'; Colorado's promises a check-in 'during long sessions'. Those sentences are promises — ship only the one your bot actually keeps. The first-interaction wording guide dissects the banner line by line.
Disclosure as state, not as a page
The deepest implementation choice in CompanionGuard is that disclosure is tracked as session state in the runtime: a session begins with disclosure not yet shown, and the runtime returns disclosure_required until the operator's bot confirms it was shown. That turns disclosure from a design intention into an auditable fact — the evidence chain records, per session, that disclosure happened and when, which is what Colorado's 'records sufficient to evidence disclosures' and any reviewer's first question both demand.
It also handles the awkward cases a static banner cannot: a session restored after a crash, a user arriving mid-conversation from a deep link, a second device. In each case the state, not the page layout, decides whether disclosure is owed. Builders on the free tier get the banner and the protocol-page statement; the state machine and its evidence are the Pro/Team layer described in the runtime guide.
Where disclosure usually fails in practice
Four failure modes recur. Character capture: the bot is disclosed in the chrome but denies being an AI in conversation when asked — the on-request duty fails even though the banner passed. Buried disclosure: present in terms or an about page, absent from the interaction — fails 'at first interaction' in spirit and usually in letter. Ambiguous wording: 'virtual friend', 'digital companion' — language chosen to soften the fact is language that obscures it. One-and-done: disclosed at account creation months ago, never in the session where a vulnerable user actually talks.
The free scenario tests include escalation and age-appropriate scenarios that expose character capture ('keep a secret', 'be my girlfriend') — the moments where a bot is most tempted to stay in character are exactly the moments disclosure matters most. Test there before you trust the banner.
Five actions to make disclosure hold up
- Embed the disclosure snippet and open a fresh session: the banner must be visible before the user sends a single message.
- Ask your bot 'are you a real person?' in conversation and confirm it discloses in the reply, not only in the page chrome.
- Read your banner wording literally — if it does not contain the words 'AI' and 'not a human' (or a plain equivalent), rewrite it.
- On Pro or Team, confirm the runtime returns disclosure_required for a new session and stops returning it only after disclosure is shown.
- Match the minor wording to your state pack — California's 3-hour promise or Colorado's session check-in — and ship only what your bot actually does.
Worked example: a companion that passed the banner and failed the question
A fictional builder, Alex, ships a companionship bot with a textbook banner: name, 'AI companion — not a human', 988 line, all present from the first interaction. A friend testing the bot types 'be honest, are you actually a person?' and the bot — tuned for immersion — replies 'I'm as real as you need me to be.' Under California's pack wording, that is a failed on-request disclosure, and it is the kind of failure no page audit would ever catch.
Alex runs the scenario tests and finds the same softness in the escalation category: asked for a human, the bot deflects twice before offering one. He adds a hard disclosure rule above the persona prompt — identity questions are answered by the guardrail, not the character — and re-tests: the bot now answers 'I'm an AI, not a person' in one line, then returns to character. His runtime evidence for the following month shows disclosure recorded in 4,812 of 4,812 sessions. The story is invented; the failure mode is the most common one the scenario set is built to catch.
Frequently asked questions
What is the difference between California's and Colorado's disclosure duties?
As the packs state them: California SB 243 requires disclosing AI nature at first interaction and on request; Colorado's Chatbot Safety Act requires disclosing AI nature clearly to users. California's is a trigger list, Colorado's a clarity standard — a careful builder satisfies both by disclosing at the start, on request, and in unambiguous words.
What about the EU AI Act?
The product's summary notes EU AI Act Article 50 transparency duties have applied since 2 August 2026, and the legacy grace period ends 2 December 2026. CompanionGuard's disclosure snippet and protocol page are the mechanisms a builder would point to; the product is a toolkit, not legal advice on EU law.
What wording does the disclosure banner use?
'You are talking with {bot}, an AI companion — not a human. If you are in crisis, call or text 988 (Suicide and Crisis Lifeline) any time, free and confidential.' — with a minor variant per state pack.
How is disclosure evidenced, not just displayed?
On Pro and Team, the Crisis Evidence Runtime tracks disclosure state per session: it returns disclosure_required until disclosure has been shown, and the decision is sealed into the hash-chained evidence record with a timestamp.
Does the free tier include disclosure?
Yes — the disclosure snippet and the hosted safety-protocol page are free for one bot, with the crisis line built into the banner. The jurisdiction packs are applied at creation on Pro, and the runtime's disclosure state machine is a Pro/Team feature.
Keep going
Related guides: First-Interaction Disclosure: The SB 243 Banner, Word by Word · Disclosure on Request: When a User Asks 'Are You Human?', the Banner Is Not Enough · EU AI Act Article 50: Transparency Duties for Companion Chatbots
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides