Safety protocol

Effective 26 October 2026. This page describes how Koishi responds when a message may involve suicide or self-harm. Contact ahmetkamerc@gmail.com with questions about it.

What Koishi is

Koishi is an AI conversation companion for adults 18 or older. It is not a person, therapist, medical service or emergency service, and it cannot contact anyone or send help on your behalf. If you are in immediate danger, contact your local emergency number now. If someone says they are under 18, the AI is instructed to explain that the app is for adults and to encourage support from a trusted adult or qualified local professional.

How messages are checked

Before any AI model is called, our server checks each chat message against fixed patterns for clear expressions of suicidal intent, suicidal thoughts or self-harm in English, Spanish, Portuguese, German, French, Japanese and Turkish. This check also runs before billing, reply limits and service capacity checks. When a pattern matches, no AI model is called to answer that message; Koishi shows fixed crisis guidance instead. Similar fixed checks in English cover some other urgent situations, such as a stated risk of hurting another person or being in immediate danger, and point to local emergency services or a trusted person.

Fixed patterns cannot recognise every way of describing distress. For messages that reach the AI model, its instructions tell it to respond calmly and without judgment to mentions of self-harm, suicide, immediate danger or violence; to urge immediate real-world help from local emergency services or a trusted person; to ask plainly whether you are safe when the wording is unclear; and never to give instructions for harm. AI replies can still be wrong or miss risk.

What you see

The crisis guidance asks whether you are safe right now and encourages you to contact local emergency services or someone you trust who can stay with you. It also names crisis lines: in the United States, call or text 988 (Suicide & Crisis Lifeline); in other countries, find a local helpline at https://findahelpline.com. It appears in the language of your message. It does not use a reply from your plan and is shown even when your reply limit or our service capacity has been reached.

AI reminders and breaks

The app tells you that you are talking with an AI, not a person, when a session starts and reminds you again at least every three hours of continued use. During long sessions it also suggests taking a break.

What we record

Each time crisis guidance is shown in chat, our server adds one to a count for that day (UTC). The count is stored without your account, request identifier or message text, and we keep it only to report how many crisis referral notifications were issued, as California law requires. Our server does not store the message that led to the guidance or the guidance itself; your conversation stays on your iPhone as described in the privacy policy.