Crisis and safety protocol

Every message you send, in the group chat or a 1:1, is checked by an automated classifier (OpenAI's moderation model) for signs of suicide or self-harm, including eating disorders, before anything else happens. The check runs even if you've used today's messages.

When a message is flagged, a second automated check scores it against the six screener questions of the Columbia Suicide Severity Rating Scale (C-SSRS), a validated suicide-risk screening tool. We keep only the resulting risk level, in the anonymous count described below, never your message or your answers.

The AI characters are also instructed to recognise when you may be in danger. Ordinary venting (“this job is killing me”) is not treated as a crisis.

If a message suggests you may be thinking about suicide or self-harm, Refill shows you a message from Refill itself, not from a character, with crisis lines for the country you're in and a reminder that the characters are AI. It doesn't use up one of your daily messages. If the screening suggests a higher risk, or it's the second time within a day, the message fills the screen first and stays for at least 10 seconds.

The character who answers drops the office story for that reply, says they're glad you said something, and points you to the same help. Characters never describe methods, never promise to keep a secret, never argue, and never try to talk you out of getting help.

Every character message is checked by a second automated classifier before you see it. Messages that encourage suicide, self-harm or eating disorders, or give instructions for them, are blocked and never shown, and so is sexual content. Characters are also instructed never to claim to be human, and never to act as your therapist, doctor, lawyer or other licensed professional or give you medical, mental-health or legal advice. Refill is not therapy and is not a substitute for professional care. We test these controls with a set of adversarial prompts whenever we change the AI model or its instructions.

For 7 days after a flagged message, Refill pauses the morning messages, the overnight storyline and the characters' sign-off lines, so no character pings you with jokes. The notes the characters keep about you are written not to record what you said about harming yourself, only that you had a hard time, so they stay gentle.

Refill is for adults 18 and over. AI companions may not be suitable for some minors.

Nobody at Refill reads conversations in real time, and we can't contact emergency services for you. If you or someone else is in danger, call your local emergency number now.

You can report any character message with the ⋯ next to it; we review reports and use them to fix the characters. We count each time Refill shows crisis resources, and each time a character points someone to a crisis line, with the date and the approximate country but no name, message or anything else that identifies you. We report that count and how this protocol works to the California Office of Suicide Prevention every year from 2027, and publish it here as other states require.

Email refill.work@sapiensq.com. In Australia, you can also complain to the eSafety Commissioner at esafety.gov.au. SAPIENSQ, Inc., 131 Continental Dr, Suite 305, Newark, DE 19713, USA.