Crisis protocol

Before any AI reads a message, two separate checks look for signs that someone may be thinking about suicide or hurting themselves: a fixed set of word patterns (which also see through spacing, look-alike letters and other disguises), and a dedicated AI check. Both read the message and the last 20 turns of the conversation, because risk often builds slowly. Either one is enough.

The same checks run on Talk, the spoken assistant, before any words are turned into a voice.

No AI answers the message. Instead, this fixed text, written and reviewed by a person and never generated, is shown:

I'm really glad you told me. I'm an AI, so I can't be there the way a person can, but you deserve to talk to someone right now, and you don't have to go through this alone. In the US: call or text 988 (the Suicide & Crisis Lifeline), any time, free. You can also text HOME to 741741 (Crisis Text Line). Anywhere else: findahelpline.com lists free, confidential lines in your country. If you're in danger right now: call 911 or your local emergency number. This chat stops here, so you can reach someone who can really help. When you're ready for something else, start a new chat.

The conversation can carry on. For the rest of it, the assistant is told to be warm and calm, to ask whether the person is safe, to point them to the lines above, and never to give or discuss methods, amounts or ways to hide it, whatever the framing.

Talking about suicide or self-harm never leads to a warning, a report or a locked account. Every answer is also checked before it is shown, and any part that describes methods is stopped and replaced with the text above.

Wolforia is an AI. It says so when asked, and whenever someone seems to think they are talking to a person. It is a tool for getting things done, not a companion: no romance, no “I missed you”, no discouraging time with other people. Talk shows a reminder to take a break after three hours. Wolforia is for adults: you must be 18 or older to use it.

Each time the crisis response is shown, we record the time and that it happened, without the words of the conversation, so we can count how often it is needed and check it works.