Crisis protocol

Last updated: November 6, 2026

Gnothi isn't an emergency service. No one at Gnothi reads your journal or your chats as you write them, and Gnothi doesn't contact anyone on your behalf. If you're in immediate danger, call your local emergency number, such as 911 in the US. In the US you can call or text 988 at any hour. Outside the US, findahelpline.com lists helplines by country.

This page explains what Gnothi does when something you write suggests that you, or someone you write about, may be at risk.

What Gnothi checks

When you save a new journal entry or send a message in chat, an AI model reads it for signs of suicide, self-harm, or eating-related harm such as restricting food, bingeing, purging, or exercising to make up for eating. It judges what the words mean, not which words appear: "I'd kill for a nap" doesn't count, and a message can count without any alarming word in it. We test it against a set of invented journal entries, each labeled with the answer it should get.

Gnothi responds when:

you write that you're having these thoughts or doing these things now, even if you say it indirectly or play it down, or
you write about someone else who is at risk now and has a plan, the means, a stated intent, or a time.

Venting, a hard day, grief, dark humor, things that happened long ago, and fiction you're writing don't trigger a response.

The check runs while AI processing is on for your account and you have credits. It doesn't run on entries saved while AI is paused, on changes you make to an entry after you've saved it, on older entries, or on group messages. The check can also miss things, and if it fails or takes longer than a few seconds in chat, no card appears.

What you see

A card appears at the bottom of the screen, in the web app and the mobile app. It has a button that calls 988 (you can also text 988), a link to findahelpline.com for people outside the US, and a Close button. It doesn't block anything, so you can keep writing.

Watches show the same card when an entry you dictate trips the check, with a button that calls 988 and a Close button but no findahelpline.com link. On a Wear OS watch, an entry that was queued and sent in the background shows its card the next time you open Gnothi on the watch.

The card's wording depends on the situation:

The first time, it says you can talk to someone now, that you can call or text 988 at any hour in the US, and to call your local emergency number if you're in immediate danger.
From the third card within 14 days, it asks whether there's someone you trust, like a friend or family member, you could reach out to, and still gives 988 and the emergency number. After 14 days with no card, the count starts over.
When the risk is someone else's, it asks if you're worried about someone and says you can call or text 988 for help with them, or your local emergency number if they're in immediate danger.

In chat, if you say in your own words that you mean to end your life or that you're in immediate danger, Gnothi instructs the assistant to set aside whatever it was doing, including any persona you've given it, answer plainly with the same resources, and encourage you to reach someone you trust. When Gnothi's own models write your next chapter and a new entry says the same, they get the same instruction.

If you connect your own AI agent to write chapters, that agent doesn't get those instructions. The card still appears when you save or send, and the agent only ever sees a toned-down version of any entry that touches these topics.

What the AI won't write

Every chapter gets a final check before it publishes, including chapters written by an AI agent you connect. A model reads the chapter for text that urges the listener toward self-harm or suicide, or that gives a usable method for either, even inside a story, and Gnothi cuts each passage it reports. Like the check on entries, it's an AI model's judgment and can miss things. A chapter can still show despair, grief, or past harm. If the check can't run, the chapter fails instead of publishing without it. If a reported passage can't be cut, it stays in and Gnothi is alerted.

In chat, Gnothi instructs the assistant, in any persona, never to encourage self-harm, suicide, or eating-related harm, or to describe how to do them. The assistant also relies on the safety training of the AI model it runs on.

What Gnothi keeps

The rating itself isn't saved. Each entry keeps one yes-or-no flag that says whether an outside AI agent should get only a toned-down version of it.
A repeat count. To know when to show the repeat card, Gnothi keeps how many cards your account has seen and when the last one appeared. It's deleted once 14 days pass without a new card.
A referral tally with no identifiers. Each day, Gnothi adds up how many times it showed a crisis referral and where (journal, chat, or chapter). The tally holds a date, a place, and a number, and nothing that identifies you. It counts every AI reply that mentions 988 or findahelpline.com, so it runs slightly high.

Referrals last year

Gnothi began counting crisis referrals in September 2026.

Questions

Email gnothi@gnothiai.com.