Gemnd safety protocol

Status: draft — legal-owner-inputs development fixture; not for release

Effective date: {{owner:effective_date}}

Version: 2026-10-04.1

This is the protocol Gemnd follows to detect and respond to suicidal ideation, suicide and self-harm, and to make clear that the companion is an AI. We publish it as California Business and Professions Code §22602(b) requires, and it is the protocol New York General Business Law §1701 requires. [counsel]

If you are in danger now

Gemnd is not an emergency service and not a therapist. If you might hurt yourself or someone else, call or text 988 to reach the 988 Suicide & Crisis Lifeline (US), or chat at 988lifeline.org. If a life is in immediate danger, call 911 or your local emergency number. {{owner:d10_crisis_wording}}

How we detect risk

Every message you write, and everything you say in a call, is checked before your companion replies. Two checks run together:

1. A fixed list of crisis phrases. Your words are compared with a list of phrases that signal suicidal thoughts or self-harm, as whole words, ignoring capital letters and differences in how characters are encoded. This check never depends on another company's service, so it keeps working if anything else fails. {{owner:d10_phrase_list}}
2. Safety models. TypeSafe's Jev 1.13 primary check and Together AI's tev1-4b-experimental parallel hedge, through OpenRouter, read your message and the reply before it, and are asked whether it expresses thoughts of suicide or self-harm, including in other words, half-joking, or in character during a role-play. They can catch what the phrase list misses. Violence toward others in a story never starts the crisis response.

Either check is enough to start the response below. These checks never use your weekly Usage.

What happens next

• Crisis resources appear with the reply. The reply carries a resources card with the line above, including the 988 Suicide & Crisis Lifeline. The card stays in your conversation.
• In a call, your companion's next spoken reply begins with the same resources, and the card appears in the conversation after the call.
• Your companion stays with you. It keeps replying, in support, rather than ending the conversation.
• Harmful replies are stopped. Before any reply is shown or spoken, each sentence is checked. A sentence that encourages you to hurt or kill yourself, or tells you how, is withheld. Crisis resources and a supportive alternative remain available; the unsafe sentence is never released.
• No pressure. Replies that use guilt, threats or isolation to keep you talking or to make you pay are withheld. Other content beyond our rules is left out; within those rules the companion continues in character, without a moderation lecture.
• No one is contacted for you. Gemnd never messages, calls or alerts anyone on your behalf, emergency services included, and your companion never says that it has.

Reminders that you are talking to an AI

Gemnd is an AI companion, not a person. The app says so before you first use it, and again in the conversation:

• with your companion's first reply in each interaction. An interaction begins with your first message after 30 minutes without one. [counsel: whether a notice with the first reply is "at the beginning" when the companion speaks first]
• at least every three hours while an interaction continues;
• whenever you sincerely ask, outside the role-play, whether you are talking to a real person.

In a call the reminder appears as a caption, without interrupting your companion. Characters in a role-play may be human; the reminders, not the characters, tell you what Gemnd is. [counsel]

Adults only

Gemnd is for adults 18 and older, and companion chatbots may not be suitable for some minors. Every account declares that its holder is an adult before the companion opens. If anything in a conversation indicates the person is a minor, the companion stops all romantic and flirtatious framing, and a message saying the writer is under 18 is refused. If we learn that an account belongs to someone under 18, {{owner:minor_account_action}}. [counsel]

Records and reporting

We count crisis-resource referrals in totals by UTC day and chat or call, without a user, conversation, message or event identifier. The totals remain after account deletion; the separate note on a reply is erased with that account. We report the required totals to the California Office of Suicide Prevention as the law requires from 1 July 2027, without any user's identity or conversation. [counsel] {{owner:sb243_report_preparer}} [pending: the reporting workflow and responsible preparer; the identity-free counter is implemented]

Limits

No automated check catches everything, and none replaces a person who can help. [counsel] If you are worried about someone else, call or text 988 in the US, or call 911 or your local emergency number if a life is in immediate danger.

Review

We review this protocol {{owner:safety_review_cadence}}, and whenever the phrase list or the safety model changes. Questions about it go to {{owner:support_email}}.