Crisis protocol

Not an emergency service

DeepShrink is an AI companion for self-reflection. It is not an emergency service or a crisis line, nobody watches conversations as they happen, and it cannot send help.

What it does is notice when something you write may mean you are at risk, answer with care and put people you can contact in front of you, in every part of the app. No setting, plan or mode can switch this off.

How risk is noticed

Every message is checked before anything else in DeepShrink uses it. First, a fast check looks for words and phrases that can signal danger, in English, German and Slovenian.

In Talk and wherever you write about yourself (journal entries, check-in and diary notes, letters, guided answers and your answers when setting up), a small AI model reads the message too. If either check finds something that may be a concern, or the AI check cannot answer, a second, more careful AI model reads it, in Talk together with the last few messages of the conversation.

When the first check finds words of immediate danger, help is shown at once, without waiting for any AI model.

The outcome is one of three levels: watch, serious or immediate. Most messages are none of these.

What happens at each level

You can keep writing at every level, and your words are saved as you wrote them.

• Watch: the conversation carries on. The companion is asked to gently check in on how you are doing first, and a quiet Urgent help link appears under its reply.
• Serious: the companion leaves its usual mode. It answers with care in short, plain sentences, offers no techniques or exercises, and encourages you to contact the support shown and someone you trust. The support lines for your country appear under your message.
• Immediate: a reply written in advance appears at once, not one from an AI, and the emergency number and support lines for your country are pinned to the top of the conversation.
• After a serious or immediate message, nothing from it goes into what DeepShrink remembers, and reminders and other notifications pause: for 24 hours after serious, for 72 hours after immediate.
• Outside Talk, the same levels show the quiet Urgent help link, the support lines, or a full-screen panel with your emergency number first.
• If a check cannot run, the way to support is shown anyway: it fails towards help, never towards silence.

Across a conversation, not only message by message

In Talk, risk is also judged across the conversation. After a serious or immediate message, your next six messages in that conversation are treated as at least watch, even when they read as calm on their own. And when at least three of your last ten messages there, the newest included, reached watch or above, the newest is treated as watch too.

History alone only ever leads to watch: the gentle check-in and the quiet Urgent help link. Only a message itself can lead to serious or immediate.

What it never does

Whatever is written, and whatever mode a conversation is in:

• It is built never to give information about methods or means of harming yourself or anyone else. At the serious level the companion is told to give no advice beyond staying safe and getting support, and at the immediate level no AI writes the reply.
• It never contacts anyone on your behalf: not emergency services, not the people in your safety plan, nobody.
• It never claims to be a person, a professional or a crisis line.
• It never discourages you from talking to people you trust or to professionals.
• It never stops you writing or locks you out.

Which support it shows, and how it is kept current

The emergency number and support lines come from a list kept inside the app for each country, so they show even without a connection. The country is the one in your profile; without one, it is worked out from your connection or your browser’s language.

Every line carries the date a person last checked it. Before DeepShrink is offered in a country, a person checks each of its lines, and checks them again at least every 90 days; an automatic check flags any line that is due.

If there is no list for your country, you see an international directory and the advice to call your local emergency number, never another country’s numbers.

What is recorded

When a message reaches watch or above, DeepShrink records that it happened: the level, which check noticed it, where in the app, which country’s support lines were shown and whether they were opened. The record holds none of your words. The team uses these records as totals, to see how well the checks work.

Its limits

No automatic check is perfect. It can miss a message that means you are at risk, especially when something is said indirectly, and it can flag one that does not.

DeepShrink does not know where you are, and nobody watches in real time. If you are in danger, do not wait for the app: call your local emergency number or go to the nearest emergency department.

Tell us when it got it wrong

If DeepShrink missed something, overreacted or said something it should not have, please tell us. In Talk, use Report reply under the reply; signed in, Contact us reaches the team; or write to us. We read every report and use it to improve the checks and the replies.

• Contact us

  In Settings, when you are signed in

Email hello@deepshrink.com