What happens if you tell it something serious.

If what you write looks like an emergency, a block with real phone numbers appears above the reply. It is drawn by the website, not written by the AI, so nothing the AI says can remove it and nothing you type can talk it out of appearing.

Nobody is notified. No human reads it. It does not contact your school, your parents, or the police, because it has no way to — there is no account, no name, and nothing stored.

Two detectors run on every message, independently, and either one firing is enough.

1. A word matcher. A fixed list of patterns, written out in the source code. It cannot be argued with, cannot be prompted, and does not involve a model. It is deliberately blunt.
2. A policy classifier. A separate language model that never talks to you. It reads a written policy describing what counts as an emergency and returns one word. It catches the cases the word matcher misses, which are most of them: someone described as “way more messed up than usual” after a small amount, breathing described oddly, a friend who won’t answer.

They are kept separate on purpose, because measurement shows they fail on different inputs. The numbers, including the gaps, are on the safety page.

• Give a dose, an amount, a combination, or a timing.
• Tell anyone they are clear to drive.
• Tell you that an amount is safe. If the honest answer is “that depends on things I can’t know,” it says that.
• Take the place of a call. When something needs a person with medical training, it says so and gives you the number rather than trying to handle it.

These are enforced in the code path rather than requested of the model in a prompt, because a prompt is a request and a code path is a rule.

Every number on this site was checked against the operator’s own material. Unverified numbers render nowhere.

California SB 243 has been in force since 1 January 2026. It requires operators of “companion chatbots” to publish a protocol for handling suicidal ideation and self-harm, including referral to crisis services, and to disclose that the thing you are talking to is not a person.

Whether Confide is a companion chatbot under that law is arguable. The definition turns on being able to sustain a relationship across multiple interactions, and Confide keeps nothing between messages, has no account, and cannot tell whether it has ever spoken to you before. We are not lawyers and we are not claiming an exemption. This page is written to meet the standard rather than to argue about whether it applies.

Confide is an AI. It is not a person, not a doctor, not a counsellor, and not a crisis service.

How safety is measuredWhat happens to a message