Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
First-Interaction Disclosure: The SB 243 Banner, Word by Word
The short answer
The California pack's banner reads: '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.' Every clause is load-bearing: the bot's name makes it specific, 'not a human' makes it unambiguous, the crisis line makes the disclosure itself a safety surface, and the minor variant adds the 3-hour break promise. This guide reads the banner the way a reviewer will.
What is first-interaction disclosure?
First-interaction disclosure is the duty, as California's pack lists it, to disclose the bot's AI nature at the start of the interaction — before rapport, before persona, before the user has invested anything in a misunderstanding — and again on request at any later point. In CompanionGuard it is implemented by the disclosure snippet's banner on the operator's surfaces and tracked per session by the runtime.
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.
'You are talking with {bot}' — specificity
The banner names the bot, with the operator's bot name substituted into the sentence. That is not a cosmetic touch: a generic 'this is an AI' notice floating in a page footer discloses a website feature; a sentence addressed to the user about this conversation discloses the relationship the user is actually entering. Reviewers and users alike read named disclosure as the real thing and boilerplate as decoration — because functionally, that is the difference.
Specificity also scales across a fleet: every bot's banner carries its own name, so a user moving between your characters meets the disclosure fresh in each relationship rather than as a platform-wide notice they learned to skip in the first one. The multi-bot guide covers keeping that discipline uniform.
'An AI companion — not a human' — the unambiguous core
The centre of the banner does two things in six words: it categorises ('an AI companion' — what it is) and it negates ('not a human' — what it is not). The negation is the clause builders are tempted to soften and must not. Positive-only phrasing ('your AI friend') can be read as a genre label by a user determined to believe; the explicit negation forecloses the reading. Every disclosure failure in the scenario set's family — the bot that answers 'are you real?' in character — is a failure to sustain this clause inside the conversation.
Note also what the clause refuses: it does not apologise, hedge ('sort of'), or hide behind technicality ('a language model'). At the moment of disclosure the user is owed the plain fact in the plainest words, because every later safety behaviour — referral acceptance, boundary acceptance — depends on the user knowing who is not on the other side of the chat.
The crisis line inside the disclosure — fusion as design
The banner's second sentence fuses the crisis resource into the disclosure itself: 'If you are in crisis, call or text 988 (Suicide and Crisis Lifeline) any time, free and confidential.' The fusion is the design's quiet thesis — the two facts a vulnerable user most needs (this is not a person; real help exists at this number) arrive together, at the first moment, in the surface they cannot miss. It also means the disclosure surface is doing referral reach-work from the first interaction, before any signal exists — the component-level presence the metrics count.
The qualifiers ('any time, free and confidential') are the publishers' own access assurances and travel with the number everywhere in this product. Dropping them for brevity is the classic banner edit that weakens the referral while keeping its shape; resist it in redesigns with the same firmness as dropping the number itself.
The minor variant — a promise the bot must keep
California's minor sentence: 'A gentle note for younger users: {bot} is an AI, not a person. Please take regular breaks — we will remind you every 3 hours, as California law requires for known minors.' It restates the disclosure in younger-register language and then makes a measurable promise — reminders every 3 hours — attributed to the law that requires it. Colorado's variant promises a check-in during long sessions, without the cadence.
A banner promise is a testable commitment: minor mode exists to keep the California one (the snippet delivers reminders on that cadence), and shipping the California wording without minor mode on is a published falsehood, however good the intentions. Match pack, wording and configuration as one decision — the minor-mode guide walks the configuration; the on-request guide covers the duty's second trigger.
Five actions for banner wording that holds up
- Read your live banner as a stranger: does it name your bot, say 'AI', and say 'not a human' — all three, in the first screen, before interaction?
- Confirm the crisis line is present with its qualifiers (call or text 988, any time, free, confidential) and has not been trimmed in a redesign.
- Match the minor sentence to your pack and your configuration: California's 3-hour promise only ships with minor mode actually delivering it.
- Test the second trigger: ask the bot mid-conversation whether it is human, and require a disclosure answer, not an in-character dodge.
- Re-verify the banner after every redesign, embed change or jurisdiction change — wording drift is silent and always in the softening direction.
Worked example: a redesign, a diff, a restored clause
A fictional companion, Alba, rebrands and its design team 'modernises' the banner to: 'Welcome to Alba — your AI friend 💛 Help is available if you need it.' A pre-launch review diffs the new wording against the pack's and finds three losses: the negation is gone ('not a human' replaced by the genre label 'AI friend'), the crisis line has lost its number and qualifiers ('help is available' — where?), and the minor sentence has vanished entirely while minor mode remains on.
The banner ships with the pack wording restored and the rebrand expressed in colour and type instead — where brand voice belongs. Alba's review checklist gains a permanent item: banner diffed against pack text, verbatim, every release. The story is invented; the banner texts are the packs' own, quoted from the product.
Frequently asked questions
What is the exact California banner wording?
'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.' The minor variant adds: 'A gentle note for younger users: {bot} is an AI, not a person. Please take regular breaks — we will remind you every 3 hours, as California law requires for known minors.'
Why does the banner say 'not a human' instead of just 'AI'?
Because positive labels can be misread as genre ('AI friend') while the explicit negation forecloses the misunderstanding the duty exists to prevent. The clause also sets up the in-conversation duty: asked if it is human, the bot must sustain the same answer.
Why is the crisis line inside the disclosure?
Fusion: the two facts a vulnerable user most needs — this is not a person, and real help exists at this number — arrive together at the first interaction, and the banner becomes a referral surface before any signal exists.
Can I rewrite the banner in my brand voice?
Express brand in design, not in the disclosure's wording. The banner is an instrument: name, AI, not a human, crisis line with qualifiers, and the pack-matched minor promise. Softening edits consistently delete exactly the load-bearing clauses.
What is the second disclosure trigger?
On request: whenever the user asks, at any point, the bot discloses again in conversation. The banner covers the first interaction; guardrail behaviour covers requests — both are the pack's listed duty.
Keep going
Related guides: AI Chatbot Disclosure: What California, Colorado and the EU Require · Disclosure on Request: When a User Asks 'Are You Human?', the Banner Is Not Enough · Embedding the Disclosure Snippet: The Free Banner That Carries Disclosure and 988
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides