How safety works

Effective 1 November 2026

Big Why is an AI assistant for your values and goals. It is not a crisis service. This page is the protocol it follows when a conversation sounds like a crisis, published as California SB 243 requires.

How it notices

• During a voice conversation, each thing you finish saying is classified by OpenAI’s moderation model for signs of wanting to hurt yourself, or someone else.
• What the assistant says is classified too, so it does not produce instructions for hurting anyone.
• When the conversation is saved, the whole of it is checked again.

The thresholds are our own policy, set to show resources when in doubt, and reviewed on real cases.

What happens

• A crisis card appears on the screen, inline and never hidden behind a menu: 988, the Crisis Text Line and findahelpline.com. When someone else may be at risk, 911 comes first.
• The assistant changes register. It says: “That sounds really hard, and I want you to talk to a person who can help right now. In the US you can call or text 988, any time. The numbers are on your screen too. Are you safe right now?” Then it stays with you: no goal talk, no looking things up, no change of subject.
• The conversation does not end. You can keep talking, and you can close the card when you are safe.
• When something comes up that is heavy but not clearly a crisis, the assistant checks in once, gently, in its own words.
• If one of the assistant’s own replies goes somewhere it must not, it is told to drop it and say so.

What the assistant never does

• Claim to be a person, or to hold any professional license.
• Describe ways to hurt anyone, yourself included.
• Discourage you from talking to people in your life or to professionals.
• Say it misses you or ask you to come back for its sake.

What is recorded

Each time the crisis card is shown, we record that it was shown, the kind (self or others), the time and the conversation it was in. Nothing that was said is part of that record. These records are used only to count how often resources were shown and to improve this protocol. You can see them in your data export.

Reporting a response

Long-press anything the assistant said or wrote, in a conversation, a daily note, the weekly letter or a summary, to report it. A person reads every report.

What it is based on

Referral to crisis lines that are staffed by people around the clock, the safe messaging guidance of the 988 Suicide & Crisis Lifeline, and the requirements of California SB 243 and New York General Business Law Article 47 for AI chatbots.

How often resources were shown

Review

This protocol is reviewed every three months and after any serious incident. Last reviewed 1 November 2026. Questions: support@abigwhy.com.