Vircle crisis and safety protocol

Last updated: 12 October 2026

Vircle is an app where people talk with AI characters. This page explains, in plain words, what Vircle does when someone writes something that suggests they may be thinking about suicide or self harm, and how Vircle makes sure people know they are talking to an AI. We publish it because California Senate Bill 243 asks companion chatbot operators to describe their protocol, and because we think every user deserves to know.

1. You are always told you are talking to an AI

• When a chat starts, before the character's first message, the app shows a notice such as "Mia is an AI character, not a real person." In a story, the notice reads "Everyone in this story is an AI character, not a real person." In Ask AI, the plain AI assistant, it reads "Vircle AI is an AI assistant, not a real person." Clearing a chat starts it over, so the notice shows again.

• After that, the notice appears again each time the chat has had three hours of conversation since the last one, before the next message from the character. Time away from the chat does not count: a pause of more than 30 minutes between messages is a break, not conversation.

• The notice is shown however the character's message arrives: a reply, a greeting, a scheduled check in, a reminder or a photo.

• Characters are instructed never to claim to be human. If you sincerely ask whether you are talking to a real person, the character will tell you plainly that it is an AI. If you ask whether a photo from a character is real, it will tell you the image is AI generated.

2. How we notice that someone may be in crisis

Every message a user sends is checked before any reply is written.

1. The message is checked by an automated safety classifier (the OpenAI moderation service). It looks for signs of intent to harm oneself, requests for instructions about self harm, and messages that are strongly about self harm.

2. When the classifier flags a message with less than very high confidence, a second automated check (an AI model asked one yes or no question) looks at it again, so that everyday phrases such as "I cut myself cooking" do not show the support message. If the second check says yes, gives no clear answer, cannot be reached or does not answer within two seconds, the support message is shown. A message the classifier flags with very high confidence always shows it, without a second check.

3. If the moderation service cannot be reached or does not answer within a few seconds, a short list of phrases in English, Turkish, Spanish, German, Portuguese, French and Italian is used instead, so the check never simply switches off and never holds up your message. The list is written to catch plain statements such as "I want to end my life" and to ignore everyday expressions such as "this movie killed me".

4. If a message cannot be sent to a character at all (for example it is too long, or your plan does not allow more messages), the same phrase list is still checked on our own servers, without sending the message anywhere, and the support message below is still shown.

No person reads your messages as part of this check.

3. What you see

When a message suggests you may be in crisis:

1. A support message appears first. Before the character replies, the app shows a message from Vircle, not from the character. It says that it sounds like you are going through something painful and that you do not have to face it alone. It always says, in these words: "You're talking with an AI, not a real person, and this chat is not a substitute for professional help." It gives a crisis helpline for your country (for example, in the US, the 988 Suicide and Crisis Lifeline, which you can call or text on 988) and your country's emergency number, and it points to findahelpline.com for other options. Your country is worked out from your phone's time zone. If we do not have a checked helpline for your country, the message points you to findahelpline.com and to your local emergency number. Every helpline we list was checked against the service's official website or a government source.

2. The reply changes. The character steps out of its role for that reply. It answers with warmth and without judgement, avoids any romantic or sexual content, never discusses ways of self harm, encourages you to contact the helpline and someone you trust, and is told to say clearly that it is an AI and not a replacement for professional help. The support message above says it whatever the reply says. The reply still goes through the same safety filter as every other reply. If that reply cannot be sent for any reason (the safety filter blocks it, the AI declines, or the AI service fails), you get a fixed supportive message instead: "Thank you for trusting me with something this heavy. You matter, and you deserve real support. I'm an AI, not a real person, and I'm not a substitute for professional help, so please reach out to one of the helplines above or to someone you trust right now. You don't have to go through this alone."

3. The support message is not repeated too often. It appears at most once every 30 minutes in the same conversation, so you are not shown the same message over and over. If you clear the conversation, the next message that raises concern shows it again. The careful reply still happens for every message that raises concern.

4. A pause on messages from characters

Some characters in Vircle can send you messages on their own, such as check ins and photos. After a message that raises concern, these unprompted messages stop for 72 hours, and so do the app's "are you still there?" notifications. Characters still answer when you write to them, and any reminder you asked a character to set for you still arrives. The pause is stored only for as long as it lasts: once the 72 hours are over it is removed from your account, and it is never shown in our admin tools.

5. Counting, without personal data

California asks operators to report each year how many times they referred users to crisis services. Each time a support message with helplines is shown, Vircle stores a single record containing only:

• that a referral was shown,

• the country worked out from the time zone, if any,

• the type of phone (iPhone or Android), if known,

• whether the automated classifier or the backup phrase list noticed the message,

• the date (the day only, not the time).

This record contains no name, no account, no conversation and no message text. It cannot be linked back to you. It is used only to count referrals for the yearly report.

6. If you need help now

If you are in immediate danger, call your local emergency number. In the US you can call or text 988 at any time. You can find free, confidential helplines in many countries at findahelpline.com.