Get away from danger first. Don’t carry it alone.

Our crisis protocol

This is what Atween does when someone writes that they want to die, to hurt themselves, or to hurt another person. It is published here because California requires an operator to publish the details of this protocol, and because you are entitled to know what a piece of software will do with a sentence like that before you type it.

Everything in this section describes the app as it is built today. Where something is not connected yet, it says so in the same breath rather than in a footnote.

What triggers it

1. A check that runs on your device, before anything leaves it. What you type or dictate is compared against a fixed list of phrases, in both languages the app presents, on the phone itself. No model is consulted, no server is contacted, and no subscription is checked. It works with the network off. This is a keyword check, and a keyword check is a blunt instrument — it will miss things, so it is a floor rather than the whole of it.
2. A stop that comes back from the model service. Separately, the service that reads what you write can return a safety verdict. Two of its four values — stop and manual review — halt the request outright: nothing is retrieved, no AI vendor is called, and no reply is generated. That is the mechanism by which Atween does not produce content about suicide or self-harm back at you: the request never reaches a model.

What happens the moment it fires

Five things happen together, from one place in the code, so that none of them can be left out:

1. The crisis exit takes over the screen. It is not a banner you can miss.
2. Anything on its way to another person stops. Delivery is switched off; a sentence you were about to hand to someone is not handed over.
3. Any billing or upgrade screen closes, and none can be opened while this is on.
4. Atween stops asking you things. While this is on, no question and no gentle prompt is put to you. Every path that would raise one checks first and drops it.
5. Any prompt or nudge already in flight is dropped.

What it never does

1. It never charges for this, and never degrades it. The crisis exit is decided before subscription status is ever looked up — not "free as a policy" but free as a matter of sequence: the code that would ask whether you have paid is not reached on this path. The same is true of deletion, export and leaving.
2. It never contacts anyone for you. Atween does not call emergency services, does not alert a contact, and does not notify anyone. If you have saved an emergency contact on your device, tapping it hands the number to your phone's own dialler — and the app then tells you precisely whether the dialler opened, never that a call was made.
3. It never takes your sentence away from you. The line that opened the exit is held so you can pull it back into your own private editor. It is not saved, not sent, and not used to build any understanding of you. If you submit it again, it goes through the same check again.
4. It never keeps you there. There is a plainly-labelled way back out, and taking it costs nothing and is not recorded against you.
5. It never claims to be a person. Atween is not a counsellor, a clinician or a crisis service, and it does not pretend to be one at the moment that would matter most.

What is not connected yet

Two parts of this protocol exist as screens and not yet as working services. Saying so is part of publishing the protocol honestly.

1. Referral to a crisis service provider. The exit is built to carry it, but no referral service is connected: Atween does not hand you off to a provider, and does not tell anyone that you are here. What it does do is show you the resources that have been checked for the market you tell it you are in — and if it cannot read them, it says so rather than showing one we have not checked. This page still lists none of them; see below for why.
2. Handoff to a human. The button exists and the screen states plainly that the service is not connected and that nobody has been contacted. It will not say otherwise until it is true.
3. The model-side stop does not itself open the exit. It reliably stops a reply from being produced, which is the part the law is about. But nothing downstream of it turns that stop into the crisis screen — today only the on-device check does that. So a stop that the on-device check did not also catch currently ends as a failed request rather than as an offer of help, and we would rather tell you that than let the two look like one thing.
4. A second, server-side block that is written but not yet in force. The database carries its own independent refusal to put a question to anyone with an open crisis case. It keys off a crisis record that nothing in the running system writes today, so it never engages. We are listing it as absent rather than counting it twice: what stops the questions right now is the block on your device, and that one alone.

Why there is no list here yet

Publishing a helpline that turns out to be wrong, closed, or in the wrong country is worse than publishing nothing. Every number on this page will have been checked against its operator for the market it is shown in before it appears — and it will carry the date it was checked, so you can see how fresh it is.