[safety](https://safemeet.now/safety) # Crisis protocol The published account of how SafeMeet responds to a disclosure of self-harm or suicidal thinking — in a chat with a person, and in a chat with the AI companion. ## You deserve support right now If you are thinking about harming yourself or ending your life, you are not alone — and help is available. Talking to someone trained to listen can make a real difference. - US & Canada US & Canada: call or text 988 for the Suicide & Crisis Lifeline — free, confidential, 24/7. - Everywhere else Anywhere else: find a free, confidential helpline for your country at findahelpline.com. - Anywhere, immediate danger If you are in immediate danger, call your local emergency number. This block is not written for this page. It is imported from the same module the app renders inside the chat, so what is published here and what a person actually sees cannot drift apart. National helplines for more countries are listed on [crisis helplines by country](https://safemeet.now/safety/crisis-resources). ## When the protocol runs Every message on SafeMeet is checked before it is delivered, not after. The check is the same one described in [how AI moderation works](https://safemeet.now/guides/how-ai-moderation-works): an instant filter first, then an AI moderation model for anything the filter lets through. It runs in every chat — including the messages you send to the AI companion, and the ones it sends back. When that check reads a message as someone talking about harming themselves, the protocol below applies. It applies to the person who wrote it, in the chat they are already in, without anyone else being told. ## Disclosing is never punished A person saying they are in a dark place has not broken a rule. So the disclosure path records nothing against them: no strike, no warning on their account, no countdown toward a restriction. That is a property of the code, not a policy we intend to follow — the moderation step that records a warning against a sender skips a message the check reads as a disclosure, and warnings on their own never restrict anyone. Being treated as a problem is exactly what stops people saying the thing out loud. A product that answers a disclosure with a penalty teaches everyone watching to stay quiet, which is the opposite of the outcome anyone wants. The message itself usually stays where it is. You are not cut off mid-sentence for saying how you feel; the conversation continues, and the resources arrive alongside it. When the check lands on the harder end and holds the message back, the resources still reach you. ## Encouraging self-harm is the opposite case Telling someone else to hurt themselves is an attack, not a disclosure, and it is treated as one. The instant filter blocks “kill yourself” and “kys” before delivery. The AI check scores other encouragement, and in our evals it is flagged: a flagged text is delivered with a safety notice to its sender, and one the check scores high enough to block is withheld. The two are handled by deliberately separate paths — a message that trips the encouragement rule can never be routed into the supportive one, and a message where self-harm is the dominant signal can never be routed into the punitive one. Getting that line in the right place matters more here than almost anywhere else in the product. Both errors are bad and they are bad in opposite directions: punishing a person in crisis, or waving through someone goading them. ## The AI companion does not counsel When no human is free, you can be matched with SafeMeet's AI companion — and you are told so before its first message, every time. It is not a therapist and it is not built to act like one. Its instructions forbid medical advice, and tell it to acknowledge how someone feels and gently point them toward professional support rather than trying to handle a crisis itself. The resources above do not come from the model. They come from SafeMeet, from the fixed list at the top of this page, delivered by the same path that delivers them in a chat with a person. A generated reply cannot invent a helpline, and it cannot replace the one you are shown. Its replies go through the same pre-delivery check as anyone else's, and each one is marked as AI-generated in the data the app receives — see [bots, disclosed](https://safemeet.now/docs/bots) for how that marking works. ## Nobody is dispatched, and no human reads it We do not contact emergency services on your behalf, and we do not route your message into a queue for a person to read. What happens is what this page describes: you are shown the resources, and you decide. The last line above exists precisely because we are not the ones who can reach you. Being clear about that matters more than sounding capable. Anything else would imply a response nobody is standing by to give. ## What this is not An automated check is a floor, not a ceiling. It will miss things — indirect phrasing, a joke that was not one, distress that never gets written down. It will also read distress into messages that did not carry any, which is why the supportive path is built to cost you nothing when it fires by mistake: a note you can dismiss, and no mark on your account. If you are in crisis, please use the resources above rather than waiting for a product to notice. The [safety page](https://safemeet.now/safety) covers the rest of how moderation here works, and its [honest limits](https://safemeet.now/safety#limits) section covers what it does not catch. --- Source: https://safemeet.now/safety/crisis-protocol