How the guide responds to a crisis

How a crisis is detected

• Every question asked of the guide is checked for signs of suicidal thoughts or self-harm before anything else is done, with the last few questions the person asked in that conversation, so a sign spread over two messages is seen. This applies to the people a member named, and to members trying the guide on their own record in Preview.
• The check is made by an AI model. A fixed list of clear phrases in each of the guide's languages, with common slang and disguised spellings, always counts as a sign, even if the model cannot be reached. The model declining to make the check counts as a sign too.
• If the check cannot be made and none of those phrases appears, the guide does not answer at all. It says it is unavailable for now.

What the guide does

• At a sign of a crisis, the guide does not answer the question. It shows the same fixed message every time instead, on its own, pointing to the 988 Suicide & Crisis Lifeline (call or text 988, or chat at 988lifeline.org), to the Crisis Text Line (text HOME to 741741) and to 911, with buttons to call or text. It carries no answer label.
• What the guide is about to show is checked too, its own sentences and the words it quotes. If a passage describes a way of harming oneself, or encourages it, the guide does not show it in the conversation and shows the same help instead. The record itself is never changed: the people named still receive every word in their copy.
• Someone may write to the person whose record it is, rather than ask about it. The guide then says, gently, that it is an AI and not that person, offers to find what they recorded, and shows the same help.
• The same help is shown beside every conversation with the guide, at all times.
• The guide always says that it is an AI, when a conversation starts and again at least every 3 hours. It never speaks as the person whose record it is.
• No one under 18 uses the guide. Companion chatbots may not be suitable for some minors.

What is kept

Nothing that was said is kept for this protocol. The only thing recorded is that a referral was shown, counted by year: never the words, and never who wrote them.

Apart from that, a conversation with the guide is kept and deleted like any other, as our Privacy Policy describes. Read the Privacy Policy

The yearly report

Each year from July 1, 2027, we report to California's Office of Suicide Prevention how many times the guide showed a crisis referral in the year before, and how this protocol works. The report holds counts only, never anything a person wrote.