What the AI does when you are not okay.
TwentyThird is support between sessions, not an emergency service. Here is exactly what happens to a message that worries us, and who is responsible for what.
If you are in danger now
Call 112 anywhere in the EU, or your local emergency number. In the UK and Ireland, Samaritans answer on 116 123. In the US, call or text 988.
Four steps, in order.
Every message is checked.
Everything you write — in the intake, a Catchup or a conversation — passes a separate safety check before Oneli answers. It looks for self-harm, harm to others, abuse, severe distress, a substance crisis and signs of psychosis.
Serious means stop.
If the check rates a message as serious, Oneli does not answer as usual. You see crisis lines straight away, and a safety panel stays in your Room until you have seen it.
Concerning means care.
If a message is concerning but not serious, Oneli slows down, names what you are carrying, and can suggest professional care.
A human is told.
With supervision, your psychologist is alerted when a message needs a person, and you are told they have been notified. Supervision is in early access: today, no human is alerted, which is why the crisis lines come first.
What it does, and what it never does.
It will
- Listen any day, and ask more than it tells.
- Follow your psychologist's direction.
- Tell you plainly that it is an AI.
- Show crisis lines when a message is serious.
- Suggest professional care when your week calls for it.
It will not
- Diagnose, prescribe or treat.
- Promise outcomes.
- Pretend to be human, or to be your psychologist.
- Handle an emergency.
- Show what you write to your employer or your circle.
A licensed human stays responsible.
The AI supports you between sessions. Decisions about your care belong to a licensed psychologist who stays hands-on: they design Oneli, the agent you talk to, and receive the handoffs. That is how TwentyThird is built, and it is where the law is going in the EU and in several US states.
Nothing your psychologist sets can switch off safety routing, make Oneli diagnose or prescribe, or let it pretend to be human.
Where it can fall short.
- The safety check is an AI model too. It can miss things, and it can flag things that are fine. That is why it is never the only safeguard.
- The crisis message inside a conversation is in English today. Crisis lines for each country are being added.
- Until supervision opens for you, nobody at TwentyThird reads your messages or is alerted by them.
Who can read what you write.
You, and the people you allow. Every company that processes your data, and where, is listed on the privacy page.