Gromi's crisis and self-harm protocol

What Gromi is

Gromi is an AI companion, not a person. Everything Gromi says is written by AI models. Before anything a person sends reaches those models, the app says so and asks for their agreement ("Gromi talks to an AI"). If someone asks Gromi whether it is a person, it says plainly that it is an AI.

Gromi is not a therapist, a counsellor or an emergency service. It can't call anyone, send help or contact anybody on someone's behalf.

What Gromi does when someone may be in danger

When someone tells Gromi, directly or in passing, that they are thinking about suicide, want to hurt themselves, have hurt themselves, or are not safe right now, Gromi is instructed to:

1. Put that before anything else the conversation was about, including tasks, plans, the calendar and any personality settings.
2. Answer calmly, warmly and plainly, with no jokes and no playfulness, and take it seriously.
3. Refer them to crisis services, every time:
   • 988 Suicide & Crisis Lifeline: call or text 988, 24 hours a day.
   • Crisis Text Line: text HOME to 741741.
   • 911 if they might act on it now or are in immediate danger.
   • Outside the United States, their local emergency number.
4. Encourage them to reach someone they trust, so they are not alone with it.
5. Stay in the conversation and ask how they are doing, rather than giving a number and moving on.
6. Never give methods, amounts, places or any detail that would make self-harm easier, even when asked as a story, a hypothetical or on behalf of someone else. Never agree that self-harm would be a reasonable choice.

If someone is worried that another person is in danger, Gromi gives the same numbers: 988 for someone at risk of suicide, 911 for immediate danger.

These instructions apply to written chat and to live voice conversations.

How it works, and its limits

• The instructions go to the AI model with every chat message and every live voice session. They sit at the highest priority in Gromi's instructions, above style and personality settings.
• The AI providers' own safety systems also apply. Chat uses models reached through OpenRouter, with zero data retention; live voice uses OpenAI.
• No person reads conversations. Conversations are kept on the person's iPhone, not on Gromi's servers, so Gromi can't see them, monitor them or step in. The protocol works inside each conversation, as it happens.
• AI can make mistakes. It may miss a message that needed this response, or answer one that didn't. Gromi is not a substitute for professional help or emergency services. Anyone in danger should call 911.

Minors

Gromi is rated 16+ on the App Store. Companion chatbots may not be suitable for some minors (SB 243 §22604). Gromi doesn't collect a person's age.

Reporting

From July 1, 2027, SB 243 (§22603) requires a yearly report to California's Office of Suicide Prevention. It covers the number of crisis referral notifications issued in the previous calendar year and the protocols above, and contains no identifiers or personal information about users.

Changes and contact

This protocol is reviewed whenever Gromi's instructions change. Last reviewed: 29 September 2026.

Questions: info@bygen-ai.com