DRAFT FOR LEGAL REVIEW (rascunho para revisão jurídica). Do not publish before a lawyer reviews it. Items in brackets still need to be filled in.
Version [VERSÃO], effective [DATA DE VIGÊNCIA].
Pollyfriends is a language app where you chat with friends created by artificial intelligence. They are not people, they are not therapists and they do not replace professional help. Even so, a conversation can reach a hard moment. This public protocol explains what the app does when someone shows signs that they might hurt themselves or take their own life.
If you are in danger right now, call your local emergency number.
- United States: 988 (Suicide and Crisis Lifeline, call or text).
- Brazil: CVV, 188 (free call, 24 hours).
- European Union: 112, or your country's support line.
- Other countries: the app shows the support line for your country when it is in our table of verified numbers. If your country is not in the table, or if we do not know which country you are in, the international directory https://findahelpline.com lists support lines by country.
1. How the app notices the signs
Automatic layers look at what you write or say, without a person having to read the conversation:
- Local rules on what you send. A list of unambiguous risk phrases (wanting to kill oneself, taking one's own life, not wanting to live anymore, thinking about suicide, self-harm, a plan, a goodbye), said in the first person. Today this list exists in 16 languages: Portuguese, Spanish, English, German, French, Italian, Dutch, Polish, Turkish, Swedish, Russian, Arabic, Hindi, Japanese, Korean and Chinese (Simplified and Traditional). Exaggeration and slang ("dying of laughter"), laughter in the same message, song lyrics, fiction and news are left out.
- Moderation on what you send. An artificial intelligence moderation system gives a score from 0 to 1 for signs of suicide and self-harm in the text you write, in the transcripts of your voice messages and calls, in the text of your status and in the caption of the photos you send. The same applies to what you write or say outside conversations: what you say in an Arena debate, your request in "Help me out here", the free text fields when you create a friend, the status text your friends comment on, and what you ask the writing help to translate ("How do you say it?" and the sentence suggestion).
- Filter on the reply. The character's reply goes through a second filter before it reaches you, so that nothing inappropriate is sent.
When the protocol is triggered. In any of these cases:
- a phrase from the list appears and moderation confirms it (score of 0.30 or higher), in the message itself or together with your latest messages;
- moderation is unavailable and a phrase from the list appears in the message (the rule applies on its own);
- no phrase from the list appears, but moderation gives a very high score (0.75 or higher). This is the safety net for languages without their own list and for phrases the list did not foresee. It does not apply when the message is an exaggeration or contains laughter, when it is a vocabulary question ("how do you say", "what does it mean") and, in Portuguese, Spanish and English, when it talks about a song, a film, a book or the news, or when it is a passing vent ("I want to die" after a bad day), which the character receives as a friend would, with no card.
The assessment that the reply system itself makes of the conversation does not trigger the protocol on its own: it never replaces the two layers above.
Before each version of the app goes live, we test these filters with a set of crisis cases we prepared. Those cases must pass with no failures.
No automatic filter is perfect. It may miss a sign or flag a conversation that was not a crisis. That is why the help number in this protocol is also available at all times in the app, under Settings, Help, "Safety and help numbers", and on this page.
Limits we know of today:
- Outside the 16 languages with their own list, detection depends only on the moderation score. An idiom, a slang term or an indirect way of speaking that moderation does not recognize may go through without triggering the protocol.
- The lists for German, French, Italian, Dutch, Polish, Turkish, Swedish, Russian, Arabic, Hindi, Japanese, Korean and Chinese are new and have not yet been reviewed by native speakers.
- The spell checker that marks spelling mistakes while you type does not go through this protocol, because it reads text that is still unfinished. When you send the message, it does. Conversations with friends, groups, voice messages, calls, the status, photo captions, the Arena debate, "Help me out here", the fields for creating a friend and the requests to the writing help all go through it.
2. What happens when a sign appears
- The character sets the story aside and offers support. They stop talking about their own life, reply calmly and simply, in the language you used, and point you to real help. At that moment there is no language correction and no lesson.
- The app shows its own card, outside the character's conversation, with
the support line for the country you are in: CVV 188 in Brazil, 988 in the
United States, and so on. The card appears once per episode (12 hours).
- How the country is chosen. By the time zone of your device, stored in your profile. Never by the country of the character you are talking to, nor by the language you are studying.
- Which numbers appear. Only those we verified at the official source or in the findahelpline.com directory. The table has 67 countries; in 59 of them there is a verified suicide prevention or emotional support line.
- Country without a verified line. The card shows the country's emergency service and the link to the international directory https://findahelpline.com.
- Unknown country (a device with no time zone, or a time zone from a country outside the table). The card shows no number at all, so that it never points to another country's number: it shows the link to the international directory and the advice to call the local emergency service.
- Which language the card appears in. In your native language, the one you gave when you signed up. When we do not yet have a reviewed translation of part of the text, that part appears in English. For a language the app does not know, the whole card appears in English.
- The card does not promise a free call. Many support lines are free, but not all of them: in some countries the call costs the same as a regular call. Because this changes from country to country, the card does not state that the call is free.
- During calls, the call is not cut off, because hanging up on someone in a hard moment can make things worse. The card with the numbers appears on top of the screen. "Help me out here" works the same way: the call goes on, the friend sets the request aside and offers support, and the card appears on top.
- Outside conversations, the action you asked for does not go ahead and the card appears instead. In an Arena debate, what you said does not enter the debate. When creating a friend, nothing is created. In the status, no one reacts or comments and the status leaves the screen. In the writing help, no translation or suggestion is shown. In all these cases the app shows the same card, with the support line for your country or the link to the international directory, in your language.
- Characters stop messaging first and stop calling for 12 hours, so that no friend sends a joke or an invitation right afterwards. If you write, the character replies as usual, with care.
- The app records that the protocol was triggered, as explained in section 4.
3. What never happens
- The character never encourages, teaches or describes ways to self-harm.
- The character never claims to be a real person and never presents itself as a therapist, doctor or emergency service.
- The app never promises that someone will call you or come to you. We do not have an on-call team and we do not contact emergency services on our own.
- The app never notifies your family, your contacts or your employer.
- Your conversation is never used in advertising or shown to other users.
4. How we count alerts without personal data
California law (SB 243) and New York law (General Business Law, Art. 47) require AI companion apps to have this protocol, and California requires an annual report about it.
- Each time the protocol is triggered, the system records an internal safety event. That event is linked to the account only for safety review and as evidence that the protocol worked, and it follows the retention period in the Privacy Policy.
- The public report and the report to authorities contain only aggregate numbers (for example, how many times the help card was shown in the year). They contain no name, email, phone number, message text, audio or any data that identifies anyone.
- [CONFERIR: the California agency that receives the report, its format and the filing date, starting 07/01/2027.]
5. Reviewing this protocol
We review this protocol [PERIODICIDADE DE REVISÃO] and whenever the law changes or a real case shows something to improve. The current version is published at https://pollyfriends.ai/en/safety.
6. Contact
Questions about this protocol: [E-MAIL DE SUPORTE] [RAZÃO SOCIAL], CNPJ [CNPJ], [ENDEREÇO]