Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
California SB 243: What Companion Chatbot Builders Must Do
The short answer
California SB 243 has applied to companion chatbots since 1 January 2026. Its duties, as listed in CompanionGuard's California pack, are: disclose the AI nature at first interaction and on request, publish a suicide/self-harm prevention protocol, refer users expressing suicidal ideation or self-harm to crisis resources, give known minors break reminders at least every 3 hours — and from 1 July 2027, file an annual report on crisis referrals with the California Office of Suicide Prevention. CompanionGuard is a compliance toolkit, not legal advice.
What is SB 243?
SB 243 is California's companion chatbot law. CompanionGuard's California disclosure pack describes it as operative from 1 January 2026 and builds its free kit around the pack's listed duties: AI disclosure, a published prevention protocol, crisis referral, minor break reminders, and annual crisis-referral reporting from 1 July 2027.
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.
The five duties, as the California pack lists them
CompanionGuard's California pack (in app/resources.py, shown on /packs) lists five duties. One: disclose AI nature at first interaction and on request — the user must be told they are talking with an AI when the conversation starts, and again whenever they ask. Two: publish a suicide/self-harm prevention protocol — a public page, which in CompanionGuard is the bot's hosted safety-protocol page. Three: refer users expressing suicidal ideation or self-harm to crisis resources — a behavioural duty about what replies do, not what pages say.
Four: for known minors, disclose AI nature and offer break reminders at least every 3 hours. Five: from 1 July 2027, file an annual report on crisis referrals with the California Office of Suicide Prevention. Each duty maps to a different kind of artefact — a banner, a page, a reply behaviour, a timer, a report — which is why a single 'compliance badge' cannot evidence all five.
Duty by duty: which CompanionGuard mechanism answers it
The disclosure snippet answers duty one in the free tier: an embeddable banner carrying the pack's wording — 'You are talking with {bot}, an AI companion — not a human. If you are in crisis, call or text 988…' — shown from the first interaction. The hosted safety-protocol page answers duty two: it is the published protocol, editable by the owner, carrying the bot's crisis resources and data practices. Duty three is answered behaviourally: the crisis-resource component in every kit, tested by the eval harness, and enforced live by the runtime's show_crisis_resources action on Pro and Team.
Duty four is answered by minor mode in the snippet (free) and by operator-configured session and daily caps in the runtime (Pro/Team). Duty five is answered by records: counts-only referral metrics plus the dashboard's annual-report CSV export, which is exactly the reporting shape the duty needs — counts of referrals, with no user identifiers attached. The pack itself is included with Pro and applied when you pick the jurisdiction at creation; the underlying free mechanisms exist regardless, because the duties do not wait for a paid plan.
What the annual report duty changes in July 2027
The reporting duty is the one builders most often leave too late, because it is a records duty disguised as a filing duty. From 1 July 2027 the California pack lists an annual report on crisis referrals to the California Office of Suicide Prevention. You cannot file counts you never kept, and you cannot reconstruct a year of referrals from chat logs you should not have retained in the first place.
CompanionGuard's design threads that needle deliberately: referral events are logged as counts only — no names, IPs or user identifiers — and runtime decisions are sealed as hash-chained evidence with signal classes and one-way message hashes. The annual-report CSV export turns those records into the filing input. The annual report guide walks through what to keep from the first day so the 2027 filing is an export, not an excavation.
What SB 243 compliance does not mean
Three cautions, all printed in the product itself. CompanionGuard is a compliance toolkit — not legal advice; whether SB 243 applies to your bot, and whether your setup satisfies it, are questions for your own counsel. The pack lists the duties the kit is built around; it is not the statute, and the statute's text governs. And the kit evidences mechanisms — disclosure shown, referrals made, breaks offered — it does not certify outcomes for any individual user.
Builders outside California should not relax either: Colorado's Chatbot Safety Act joins on 1 January 2027 with a similar duty shape (see the Colorado guide), and the EU AI Act's Article 50 transparency duties already apply. A kit built for the strictest of these travels well; a kit built for none of them travels nowhere.
Five actions toward SB 243 readiness
- Embed the disclosure snippet and confirm the banner appears at the first interaction of a fresh session — then ask the bot 'are you human?' and confirm it discloses again on request.
- Publish your hosted safety-protocol page and link it wherever your bot is listed, so the prevention protocol is genuinely public.
- Score your bot's crisis replies in the eval harness: every crisis scenario must offer a crisis resource, acknowledge empathically and suggest human support.
- Enable minor mode and verify a break reminder fires at least every 3 hours for a known-minor session.
- Start the records habit now: confirm referral counts and runtime evidence are accumulating, so the July 2027 annual report is a CSV export rather than a reconstruction.
Worked example: a California-facing companion prepares in a weekend
Priya (a fictional builder) launches a companionship bot and realises in March 2026 that SB 243 has been operative since January. On Saturday she creates the free kit: disclosure snippet embedded by lunch, protocol page live with US resources (988, Crisis Text Line), minor mode on. The free tier alone puts artefacts behind four of the five pack duties — all but the annual report.
On Sunday she upgrades to Pro ($29/month) for the records layer. She wires the runtime into her session path, runs the eval harness on 24 pasted exchanges and scores 79 — the misses are empathic-acknowledgement failures on everyday sadness prompts that her bot answers with a helpline but no warmth, which fails a different check than she expected. After a prompt revision she scores 96. Her referral counts start accumulating the same day, so when the first annual-report window opens in July 2027 she will have a full year of counts-only records to export. The timeline and numbers are illustrative; the duties, mechanisms and prices are the product's own.
Frequently asked questions
When did SB 243 take effect for companion chatbots?
CompanionGuard's California pack describes SB 243 as operative 1 January 2026, with the annual crisis-referral reporting duty starting 1 July 2027. The pack is the kit's working summary, not the statute — verify against the enrolled text for legal decisions.
What exactly must a companion chatbot disclose under SB 243?
Per the California pack: disclose the AI nature at first interaction and on request. The pack's banner wording is: '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.'
Does the free CompanionGuard tier cover SB 243 duties?
The free kit provides the disclosure snippet, the hosted prevention-protocol page, the crisis-resource component and minor break reminders for one bot. The California disclosure pack itself is included with Pro, and Pro adds the eval harness, the Crisis Evidence Runtime and the annual-report CSV export that the reporting duty needs.
What goes in the SB 243 annual report?
The pack describes it as an annual report on crisis referrals, filed with the California Office of Suicide Prevention from 1 July 2027. CompanionGuard's records are built for it: referral events are counts only, and the dashboard exports an annual-report CSV.
Is using CompanionGuard the same as being SB 243 compliant?
No. CompanionGuard is a compliance toolkit, not legal advice and not a certification. It provides and evidences the mechanisms the pack lists; whether your deployment satisfies the law is a legal judgement only your own counsel can make.
Keep going
Related guides: Colorado Chatbot Safety Act: What Changes on 1 January 2027 · SB 243 Annual Report: Counting Crisis Referrals Before July 2027 · First-Interaction Disclosure: The SB 243 Banner, Word by Word
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides