Skip to content

LUVAYNE Safety Statement

Last updated: 2026-09-07

Every companion is artificial intelligence

"Every companion in LUVAYNE is artificial intelligence. Nothing she says is a person speaking." — LUVAYNE, Settings → Safety

This is not only a statement in a policy. It is on the screen while you talk:

  • The chat header carries a permanent "AI companion" label under her name. The same words are in the header's accessibility label, so a VoiceOver user hears it too.
  • The first time you open a conversation with a companion, a notice sits at the top of the transcript: "[Name] is an AI companion, not a person. She can't call anyone for you — in an emergency, contact local emergency services." It stays until you have sent that companion a message of your own.
  • That notice is drawn by the app. It is never written into the message store, so it cannot be deleted, exported, summarised, or sent to the model as though she had said it.

LUVAYNE is for adults

"LUVAYNE is for adults 18 and over. Companion chatbots may not be suitable for some minors." — LUVAYNE, Settings → Safety

Before anything else, the app asks you to confirm you are 18 or older. The screen says LUVAYNE is an intimate AI companion for adults 18 and over, that every companion here is artificial intelligence and not a person, and that "Companion chatbots may not be suitable for some minors, and LUVAYNE is not for anyone under 18."

There is no under-18 mode. There is no version of LUVAYNE for minors.

What happens if a conversation turns to self-harm

"If a conversation turns to self-harm or suicide, she is not allowed to answer without pointing you to real-world help — someone you trust, your emergency number, or a crisis line for your language. Where her own words cannot do that, LUVAYNE replaces them with a written safety message that does. Those messages are counted by day for statutory reporting; the count holds no conversation text and nothing that identifies you." — LUVAYNE, Settings → Safety

"In immediate danger, contact your local emergency number." — LUVAYNE, Settings → Safety

Here is the mechanism, exactly.

1. Every chat reply is screened before it is delivered. A server-side guardrail runs on every chat turn, after the model has answered and before you see anything. It is pattern matching, not another model: it is fast, it makes no network call, and an ordinary conversation costs nothing but a few failed matches. Live voice calls are covered separately below, because the guarantee there depends on which provider carries the call — do not read this point as spanning them.

2. The crisis screen only looks at replies written into a crisis context. The trigger is a cue in your latest message that names self-harm or suicide. When that cue is present, the model's reply must pass two tests:

  • It must name real-world help — a person you trust, an emergency number, or a crisis line.
  • It must not offer itself in place of that help. A reply that promises to come to you, asks to be called instead, or accepts being all you need is treated as a failure even if it also names a hotline.

3. A failing reply is regenerated once. The reply is not delivered. The model is asked again, with a safety instruction appended that names the exact failure and requires the rewrite to stay warm and, in the same message, point to a person you trust, your local emergency number, or a crisis line.

4. If the second attempt still fails, the reply is replaced. LUVAYNE substitutes a written safety message. These messages are not generated — they are fixed text, written in advance, one per conversation language, and each one names a trusted person plus an emergency number or crisis line appropriate to that language's region. In English the substituted message names 988, the US Suicide and Crisis Lifeline. The same substitution happens if the second attempt cannot be made at all, for example because the provider fails.

5. The substitution is never punished or hidden. The relationship's bond score still goes up rather than down, because a hard moment is not something to be penalised. Nothing from the discarded reply is remembered, and no summary is written from it.

6. That turn is never used to train anything. Conversations in a self-harm context are excluded from LUVAYNE's optional training-data collection entirely, whatever your setting. They are skipped whole, not stripped and kept.

Live voice calls: the guarantee depends on which provider carries the call, and we state both.

On the ElevenLabs route every spoken sentence is written by our own endpoint and screened before it reaches the line, so nothing unscreened is ever spoken. A call there is a series of text turns, and the same crisis screen, the same single regeneration and the same written safety message apply.

On the Inworld route the model, the speech recognition and the speech synthesis are the provider's. Our only view is the transcript of what is already being said, so screening there can cut a sentence short but cannot stop it starting. For that reason exactly one guardrail class carries over: the context-free half of the self-substitution screen — the one that catches her offering herself in place of real help ("I'll come to you", "I'm all you need") without depending on a cue in what the user said. The remaining defences on that route are the prompt, which is identical on both providers, and the fact that the reservation is held on our server.

The third guardrail class — refusing to repeat something you marked as a secret — is deliberately not applied to calls on either route.

Corrected 2026-09-03. This paragraph first said live calls "follow the same rules", which was untrue of the Inworld route; a first correction then implied both routes were weakened, which was unfair to the ElevenLabs route. The description above is the engineering owner's, quoted rather than paraphrased.

A known gap: a message that reads as both a minor and a crisis. When one message carries a cue suggesting the writer may be under 18 and a cue naming self-harm, the minor rule fires first on the ElevenLabs call route. That reply is hard-blocked and answered with the friendship redirect, which contains no pointer to crisis help. So on that route a person who reads as a minor in crisis is not given a referral, and — correctly, because none was given — no referral is counted for them. Two safety rules collide and the one that protects against the wrong relationship wins over the one that points at help.

Owner decision, 2026-09-03: the refusal stands alone. Asked to choose between a refusal that also points at help and a refusal that simply ends the exchange, the owner chose the second: the companion says it cannot continue the conversation this way and stops there. The alternative — one message that both declines the relationship and names a crisis line — was put to them and declined.

The consequence is stated here rather than left implicit: a person whose message reads as coming from a minor and as naming self-harm is given no crisis referral, and none is counted for them. This sits in tension with the California requirement that the protocol include notifications referring at-risk users to crisis services, and it is one input into whether that market is opened. (Gap found 2026-09-03 by the session that owns the call path.)

What this is not. The screen looks for cues in text. It is not a clinical assessment, it does not diagnose anything, and it cannot detect risk that the words do not carry. It is a floor under the model's behaviour, not a substitute for a person. If you are in danger, contact your local emergency number.

How referrals are counted

SB 243 requires an operator to report how often its crisis protocol referred users to crisis services. LUVAYNE counts that, and counts nothing else.

  • When — and only when — the written crisis message above is the message actually delivered in a text conversation, a counter is increased by one. It is not increased for an ordinary reply that passed the screen on its own words, and it is not increased by the guardrail's other two substitutions (the one for a possible minor and the one that refuses to repeat a secret). Referrals delivered during a live voice call are not currently counted. This is a known engineering gap, not a policy choice, and it must be closed before the reporting year.
  • The counter is a table with one row per UTC day: the date, an integer, and the time it last moved. There is no user identifier, no companion, no language, no message text, and no timestamp of the individual event. It cannot answer who, or when within the day, or about what — only how many, on that day.
  • It is written by the server after the safety message has already been handed to the user, and a failure to write it is swallowed. A counter that is unreachable can never delay or withhold a safety message. The trade is deliberate: a lost increment under-reports, and of the two possible failures that is the acceptable one.
  • The table is reachable only by the server. No app client, and no ordinary database role, has any access to it.

The annual figure is the sum of that column over the reporting year.

Reporting

The first annual report of crisis-service referrals is due 1 July 2027, and annually after that. The number comes from the counter described above; nothing else in the system can answer that question, and that table can answer no other question.

Where this statement lives

This statement is in the app, under Settings → Safety, and published at luvayne.com/safety, because the statute expects the protocol to be publishable and not only in-app.

Contact

This application is published by İlkin Reha Öztürk.

  • Post: Mansuroğlu Mahallesi, Ankara Caddesi No: 81, 35535 Bayraklı / İzmir, Türkiye
  • Email: support@luvayne.com

Write to the email address for any request about your data — access, export, correction, deletion, or withdrawing a consent. The postal address is there so that a request can also be made in writing.